Next Article in Journal
Performance of Bond between Old and New Concrete Layers: The Effective Factors, Durability and Measurement Tests—A Review
Previous Article in Journal
Advances in Steel and Composite Steel—Concrete Bridges and Buildings
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Developing a Machine-Learning-Based Automatic Incident Detection System for Traffic Safety: Promises and Limitations

by
Osama ElSahly
* and
Akmal Abdelfatah
College of Engineering, Department of Civil Engineering, American University of Sharjah, Sharjah P.O. Box 26666, United Arab Emirates
*
Author to whom correspondence should be addressed.
Infrastructures 2024, 9(10), 170; https://doi.org/10.3390/infrastructures9100170
Submission received: 18 August 2024 / Revised: 23 September 2024 / Accepted: 24 September 2024 / Published: 26 September 2024

Abstract

:
This study presents a novel, machine-learning-based Automatic Incident Detection (AID) system for freeways. Through a comprehensive analysis of existing AID systems, the paper identifies their limitations and key performance metrics. VISSIM, a traffic simulation software, is employed to generate diverse, realistic traffic data incorporating factors significantly impacting AID performance. The developed system utilizes an Artificial Neural Network (ANN) trained via RapidMiner software. The ANN is designed to learn and differentiate normal and incident traffic patterns. Training yields a Detection Rate (DR) of 95.6%, a False Alarm Rate (FAR) of 1.01%, and a Mean Time to Detection (MTTD) of 0.89 min. Testing demonstrates continued effectiveness with a DR of 100%, a FAR of 1.29%, and a MTTD of 1.6 min. Furthermore, a sensitivity analysis is conducted to assess the influence of individual factors on system performance. Based on these findings, recommendations for enhancing AID systems are provided, promoting improved traffic safety and incident management. This research empowers transportation authorities with valuable insights to implement effective incident detection strategies, ultimately contributing to safer and more efficient freeways.

1. Introduction

Transportation forms the backbone of our daily lives, enabling the seamless movement of people and goods. As this demand intensifies, so does the frequency of traffic incidents. These unforeseen disruptions, caused by breakdowns, collisions, weather events, or roadworks, pose significant challenges [1,2,3,4,5]. Beyond the immediate inconvenience, traffic incidents hold grave consequences [4,6]. According to the World Health Organization (WHO), they are a leading cause of death globally, claiming millions of lives annually [7,8]. The economic impact is equally staggering, with incidents costing an estimated 3% of global GDP [7]. Their environmental toll is substantial as well, with congestion leading to increased air and noise pollution, along with greenhouse gas emissions. In the pursuit of sustainable cities, traffic incidents act as a major hurdle, hindering progress across the social, economic, and environmental pillars of sustainability. To mitigate these negative impacts, considerable efforts have been directed towards incident management. Automatic Incident Detection (AID) systems stand as a crucial tool in this fight. By automatically and accurately detecting incidents in real-time, AID systems facilitate a timely response. This allows for the rapid deployment of emergency services, the swift resolution of incidents, and a faster return to normal traffic flow. Studies have revealed a strong correlation between the duration of incidents and the potential for secondary incidents. As an incident continues to unfold, the likelihood of additional crashes occurring within the congested area increases, further exacerbating the situation [9,10,11,12,13,14]. Furthermore, traffic incidents are estimated to be responsible for a significant portion of traffic delays [14], underscoring their disruptive effect on traffic flow. In this context, AID systems play a vital role in modern traffic management. By enabling prompt incident identification and response, they contribute to enhanced road safety, reduced casualties, mitigated economic losses, and improved environmental sustainability [15,16,17]. Recognizing their importance, researchers have continuously striven to develop effective and efficient AID solutions. This paper builds upon these efforts by proposing a novel AID model. The model aims to achieve high accuracy and prompt detection of traffic incidents while maintaining robustness against variations in incident patterns. To achieve this goal, the paper will delve into existing AID systems, analyzing their functionalities, strengths, and limitations. Subsequently, it will explore the factors influencing AID system performance, as understanding these factors is essential for designing a robust and effective model. Finally, the paper will introduce and evaluate the proposed AID model, demonstrating its ability to overcome limitations of existing approaches and effectively detect incidents in diverse scenarios. Through the presentation of this novel model, the paper aspires to contribute to the ongoing advancements in smart city development by offering a valuable solution for enhanced traffic management.

2. Literature Review

The development of Automatic Incident Detection (AID) systems has been a significant area of research since the 1970s [18,19,20,21,22,23]. Early methods relied on non-automatic approaches, such as eyewitness reports, which were subjective and time-consuming [24]. Modern AID systems leverage advanced technologies to automatically collect and analyze traffic data, improving the accuracy and timeliness of incident detection [25,26,27,28,29,30,31,32,33].
The core principle underlying AID systems lies in the recognition that traffic incidents often cause lane blockages, leading to bottlenecks at specific locations [34]. This disruption manifests as a discontinuity in traffic flow, resulting in observable variations between the upstream and downstream sections of the affected road. These variations typically include reduced speed and volume upstream coupled with increased occupancy, while downstream experiences elevated speed and reduced volume/occupancy [34,35,36,37,38,39].
The landscape of AID systems is diverse, exhibiting a wide range of data processing methods and detection algorithms. These systems can be broadly categorized into four main types: comparative models, statistical models, image processing models, and artificial-intelligence-based models [40,41,42]. Each of these categories adopts a distinct approach to incident detection, relying on different computational strategies and methodologies. The following sections of this literature review will provide a comprehensive examination of each AID category. The analysis will delve into the main performance evaluation measures, distinguishing features, advantages, and drawbacks associated with each type of system. By synthesizing the knowledge accumulated since the inception of AID systems, this review aims to offer a thorough understanding of the characteristics and capabilities of the diverse spectrum of models employed for automatic incident detection.

2.1. Comparative Incident Detection Algorithms

Within the diverse landscape of AID systems, comparative algorithms occupy a prominent position. These algorithms detect incidents by comparing traffic parameters such as speed, volume, and occupancy between different sections of a roadway. Significant discrepancies between upstream and downstream data suggest the presence of an incident [25,40,41]. Several well-established algorithms illustrate the comparative approach:
  • California Algorithm: Detects incidents based on differences in occupancy between adjacent detectors, indicating possible congestion due to an incident [21,22,34,41,43,44,45,46,47,48,49,50].
  • Pattern Recognition (PATREG): Expanding on the California algorithm, PATREG incorporates historical traffic patterns into its detection logic [51]. By comparing current data with established patterns, it identifies deviations that might signal an incident. However, its dependence on historical data can limit its adaptability to dynamic traffic conditions or novel incident types.
  • All-Purpose Incident Detection (APID): Utilizes multiple detection routines tailored for various traffic conditions, incorporating additional tests such as compression wave and persistence tests to enhance accuracy [52].
While offering valuable tools for incident detection, comparative algorithms come with inherent limitations:
  • Susceptibility to False Alarms: External factors such as weather, lane closures, or sudden traffic volume changes can trigger false alarms, leading to resource misallocation and response delays [41,47,53].
  • Limited Adaptability: Algorithms relying solely on predefined thresholds or historical patterns might struggle to adapt to dynamic traffic conditions or novel incident types [46].
  • Data Dependence: Their performance heavily relies on the quality and accuracy of input data from detectors or other sources. Issues with data collection or transmission can negatively impact their detection accuracy [46].

2.2. Statistical AID Algorithms

These algorithms establish normal traffic patterns using statistical models and detect deviations from these norms that might indicate an incident. These algorithms apply statistical tests or metrics to traffic data to identify unusual behavior [30,54,55,56].
Several prominent algorithms exemplify the statistical approach:
  • Standard Normal Deviate (SND) Algorithm: Detects incidents by calculating the standard deviation of traffic parameters and flagging significant deviations from the norm [19,41,57].
  • Bayesian Algorithms: These algorithms employ Bayesian statistics to continuously update the probability of an incident based on incoming traffic data [24,43]. They offer flexibility in incorporating prior knowledge and adapting to changing conditions but require careful model design and parameter selection.
  • High Occupancy, Low Speed, and Congestion Criterion (HIOCC): Identifies potential incidents by detecting the concurrence of high occupancy, low speed, and significant congestion, surpassing predefined thresholds [51]. While robust to isolated anomalies, its dependence on multiple criteria can reduce sensitivity to certain incident types. These models analyze historical traffic data as time series, identifying patterns and trends [26,40,41,55,56,57,58,59,60]. Statistical methods such as Autoregressive Integrated Moving Average (ARIMA) can then be used to predict future traffic flow [55,56,57,58,59,60]. Significant deviations from these predictions might indicate an incident. While offering adaptability, their effectiveness relies on the quality and representativeness of historical data.
Statistical algorithms offer several advantages:
  • Adaptability to Changing Conditions: By analyzing patterns and trends, they can potentially adapt to dynamic traffic conditions better than methods relying solely on fixed thresholds.
  • Incorporation of Prior Knowledge: Bayesian approaches allow for incorporating historical data and domain knowledge, potentially improving detection accuracy.
  • Computational Efficiency: Some methods, such as SND, are computationally efficient and suitable for real-time applications.
However, limitations also exist:
  • Sensitivity to Data Quality: Their performance heavily relies on the quality and accuracy of input data [60,61,62,63].
  • Model Complexity: Complex models such as Bayesian approaches can be computationally expensive and require careful calibration [41].
  • False Alarm Potential: Deviations from normality can occur due to factors other than incidents, leading to potential false alarms [38].

2.3. AI-Based AID Models

Artificial Intelligence (AI) and machine learning (ML) algorithms have revolutionized various fields, and traffic management is no exception [64,65,66,67,68,69,70,71]. Machine learning empowers computers to learn from data without explicit programming. By analyzing vast datasets, ML algorithms can identify patterns and relationships, enabling them to make predictions or classifications on new, unseen data [61,72]. In the context of traffic incident detection, AI and ML algorithms are employed to develop AID systems [62,63,73,74,75,76,77,78]. These systems leverage historical or real-time traffic data, such as speed, volume, and occupancy, to learn the characteristics of normal traffic flow. This training process allows the ML model to differentiate between normal and abnormal traffic patterns that might indicate an incident. Once trained, the model can then classify incoming traffic data as either “normal” or “incident”, enabling real-time detection of potential disruptions.
Several ML algorithms have been successfully implemented in AID systems. Some prominent examples include:
  • Artificial Neural Networks (ANNs): Inspired by the biological structure of the brain, ANNs consist of interconnected layers of processing units that learn complex relationships within the data [79,80,81,82,83]. They excel at identifying non-linear patterns in traffic flow, making them well-suited for incident detection [5,77,84,85,86,87,88,89,90,91,92,93].
  • Random Forests (RFs): Random Forests are ensemble methods that combine multiple decision trees to improve accuracy and reduce the impact of noisy data [94,95,96,97,98]. Furthermore, they offer some level of interpretability, allowing researchers to understand which features are most important for incident detection [39,98,99,100].
  • Fuzzy Logic (FL): This technique incorporates the concept of partial truths [101,102,103,104], allowing for nuanced evaluation of traffic data that might not fall strictly into predefined categories. This flexibility can enhance the sensitivity of incident detection, especially in scenarios with ambiguous data [105,106,107,108,109].
  • Support Vector Machines (SVMs): SVMs excel at finding the optimal hyperplane that separates data points belonging to different classes (normal vs. incident traffic) [97,110,111]. Their ability to handle high-dimensional data is advantageous for analyzing complex traffic patterns [9,38,98,112,113].
  • Hybrid Models: Researchers are increasingly exploring the potential of combining different AI algorithms or AI with other techniques such as statistical methods. This fusion approach can leverage the strengths of each individual technique to create more comprehensive and robust AID models [114].
These ML-based AID systems offer several advantages:
  • Learning Ability: They can continuously learn and improve their detection accuracy with exposure to new data.
  • Real-time Processing: ML algorithms can analyze traffic data in real-time, enabling prompt incident detection.
  • Adaptability: They can be adapted to different types of roadways and traffic conditions.
However, some limitations also exist:
  • Data Dependence: The performance heavily relies on the quality and quantity of training data.
  • Computational Cost: Training complex ML models can be computationally expensive.
  • Black Box Phenomenon: Their internal workings can be opaque, making it challenging to understand and interpret their decision-making processes.
Despite these limitations, machine learning has emerged as a powerful tool for developing robust and effective AID systems. By addressing these challenges and continuously improving ML algorithms, researchers can further enhance the accuracy, efficiency, and explainability of these systems, leading to safer and more efficient traffic management.

2.4. Image Processing AID Algorithms

Leveraging the capabilities of computer vision and image processing techniques, these models analyze video data captured by surveillance cameras mounted along roadways. By dissecting these videos into individual image frames, crucial traffic variables such as vehicle volume, speed, and occupancy are extracted, enabling incident detection based on visual cues gleaned from the analyzed imagery [25,42,55,115,116,117,118,119,120,121,122].
Image Processing AID models offer numerous advantages:
  • Direct Observation: These models directly observe the traffic scene, potentially providing richer information compared to models solely relying on sensor data.
  • Versatility: They can be adapted to various camera configurations and environmental conditions, offering flexibility in deployment.
  • Identification of Specific Incidents: Analyzing visual cues enables the identification of specific types of incidents, such as car accidents or disabled vehicles, which might be challenging for other methods.
However, limitations also exist:
  • Computational Demands: Processing video data can be computationally expensive, requiring powerful hardware and optimized algorithms.
  • Weather Dependence: Visibility limitations due to rain, snow, or fog can hinder performance and lead to false alarms [121].
  • Privacy Concerns: The use of video data raises privacy concerns that require careful consideration through anonymization techniques and responsible data management practices.

2.5. Evaluating the Performance of AID Models

Evaluating AID models requires a nuanced approach due to the inherent complexities of real-world traffic environments. Incident detection is a binary classification problem, categorizing traffic states as either normal or abnormal (incident), and several key metrics are used to assess performance.
Accuracy is a basic indicator of correct classifications, but in imbalanced datasets, where incidents are less frequent, it can be misleading by masking poor performance in detecting actual incidents [123,124]. Precision measures the proportion of correctly identified incidents but may overlook genuine incidents to minimize false alarms [123,124]. Recall (or True Positive Rate) focuses on the proportion of actual incidents detected but may increase false alarms [123,124]. To balance these, the F1-score, which is the harmonic mean of precision and recall, offers a more comprehensive evaluation, particularly useful in imbalanced datasets [123,124].
Beyond binary classification metrics, AID model performance is typically evaluated using Detection Rate (DR), False Alarm Rate (FAR), and Mean Time to Detect (MTTD) [125]. DR, similar to Recall, indicates the percentage of incidents successfully detected [41,115,125]. FAR represents the proportion of non-incidents incorrectly classified as incidents [125,126,127,128], while MTTD measures the average time taken to detect an incident [126]. Balancing these metrics requires careful calibration, as increasing DR may raise FAR [41,129,130]. Thus, optimizing these metrics demands a balanced approach to ensure accurate detection while minimizing unnecessary alarms and resource misallocation.
Despite significant progress in developing AID systems, existing models still face persistent challenges, particularly in accurately and efficiently detecting incidents under varying traffic conditions. These shortcomings include limitations in adapting to diverse traffic environments, handling varying incident severities, and integrating different types of data inputs. The primary objective of this study is to address these limitations by utilizing machine learning techniques to create a more efficient and generic incident detection model. Unlike previous studies, this research simultaneously considers four critical factors that significantly affect AID system performance: congestion levels, incident severity, incident location, and the distance between detectors. By incorporating all these factors together, this study aims to produce a more realistic and adaptable solution, providing a deeper understanding of how AID systems behave in dynamic traffic conditions. Ultimately, the goal of this research is to enhance the accuracy and speed of traffic incident detection, contributing to safer and more efficient road networks. This improved performance can help reduce the economic, safety, and operational impacts of traffic incidents, making a meaningful contribution to transportation management.

3. Methodology

This chapter outlines the methodology employed in this study to develop and evaluate a novel, machine-learning-based Automatic Incident Detection (AID) system for freeways. The research focused on a specific study area (to be specified) representative of freeway traffic conditions. The following sections will detail the data generation process, system development stages, and the evaluation techniques used to assess the model’s performance.

3.1. Study Area Selection

The selected road is a 125 km section of a major existing freeway in the UAE, with a maximum speed limit of 130 km/h. This section includes six lanes in the basic segments, with lane numbers varying between six and seven throughout. It features six junctions: two right-in-right-out junctions, a single-point interchange, and two full-cloverleaf junctions with additional ramps as shown in Figure 1 below. These junctions require lane-changing and weaving maneuvers, which introduce turbulence in traffic flow, posing an added challenge for the developed model to avoid misclassifying them as traffic incidents.
To ensure the microscopic model closely replicates real-world conditions, the geometric properties of the road, including lane width, curves, the number of lanes, junction locations, vehicle movements, and posted speed limits, were accurately modeled. This approach addresses concerns regarding the simplification of vehicle movements in simulations, ensuring that the model reflects the actual conditions of the selected freeway.
To account for the worst-case scenario, road capacity calculations were based on the seven-lane segments, with a total capacity of 16,800 passenger cars per hour, following the Highway Capacity Manual (HCM) standard of 2400 cars per hour per lane [131]. By incorporating this complexity into the training data, the model is better equipped to differentiate between normal traffic flow and disruptions indicative of genuine incidents.
The selected road is a 125 km section of freeway with a maximum speed limit of 130 km/h. This section includes six lanes in the basic segments, with lane numbers varying between six and seven throughout. It features six junctions: two right-in-right-out junctions, a single-point interchange, and two full-cloverleaf junctions with additional ramps. These junctions require lane-changing and weaving maneuvers, which introduce turbulence in traffic flow, posing an added challenge for the developed model to avoid misclassifying them as traffic incidents. To account for the worst-case scenario, road capacity calculations were based on the seven-lane segments, with a total capacity of 16,800 passenger cars per hour, following the Highway Capacity Manual (HCM) standard of 2400 cars per hour per lane. By incorporating this complexity into the training data, the model would be better equipped to differentiate between normal traffic flow and disruptions indicative of genuine incidents.

3.2. Data Generation and Development of the Simulation Model

In developing traffic incident detection models, two main sources of traffic data are typically available: real-world data and simulated data. Real data, collected from sensors, cameras, and GPS devices, provides direct observations of actual traffic conditions, offering valuable insights into vehicle movements, incident occurrences, and external factors such as weather. However, real data collection involves high costs due to the need for expensive equipment and ongoing maintenance. Additionally, real data is often limited in coverage, only available in areas where sensors are installed, and can be prone to inaccuracies due to environmental factors such as weather. Access to comprehensive real data, particularly detailed information about incidents (e.g., severity, location, and exact time of occurrence), can also be challenging and restricted.
On the other hand, simulated data generated by traffic simulation software offers several key advantages. It is cost-effective, eliminating the need for extensive infrastructure, and provides flexibility by allowing the modeling of various traffic conditions and incident scenarios that may be difficult to capture using real data. Simulated data also allows for precise control over variables and scenarios, making it easier to analyze the impact of specific factors on incident detection models. However, simulated data is not a perfect substitute for real-world data. It is based on assumptions and simplifications of real-world conditions, and while highly flexible, it may not capture the full complexity of actual traffic behavior.
In this study, simulated data were selected due to their flexibility, cost-effectiveness, and ability to generate diverse traffic scenarios, which are critical for developing robust incident detection models. While real data could also be used with the proposed models, capturing a wide range of conditions and incidents through real-world data collection alone would be impractical. Additionally, several studies have demonstrated the effectiveness of using simulated data for developing AID models, supporting the decision to utilize this approach.
The data generation process specifically focused on incorporating four crucial factors that, based on the literature review, are believed to have a significant impact on the performance of AID models: traffic congestion level, incident severity, location of the incident, and the distance between traffic detectors [131]. These factors are complex and can interact with each other in intricate ways, posing a challenge for considering all of them in a single model [105]. None of the existing models identified in the literature review have been designed to consider all of these factors together. However, the model developed in this paper will address this gap by simultaneously considering these four factors, aiming to achieve superior performance and generalizability. This comprehensive approach acknowledges the complex interplay of these factors in real-world situations, paving the way for a more reliable and adaptable model.
To generate the simulated traffic data, VISSIM, a widely used microscopic traffic simulation software [132], was employed to model the selected study area. The study area is a major existing freeway located in the UAE. The geometric parameters of the study area, such as lane configurations, speed limits, and junction layouts, were carefully modeled in VISSIM to replicate the actual roadway characteristics. This ensures that the vehicle movements in the simulation closely mirror real-world conditions. VISSIM uses detailed driver behavior models, which allow for realistic representation of traffic flow and vehicle interactions.
To accurately simulate driver behavior and vehicle interactions, the Wiedemann 99 car-following model was utilized in this study. This model is designed for freeways and high-speed roads, making it particularly suitable for the selected study area. The Wiedemann 99 model simulates the behavior of individual drivers by considering parameters such as the following distance, speed differences, and driver reactions to vehicles ahead. It incorporates four driving regimes: free-flow driving, approaching, following, and emergency braking. By adjusting these parameters, the model replicates the varying behaviors of drivers, from normal cruising to abrupt braking in response to incidents.

3.3. Simulated Traffic Data Collection Parameters

VISSIM, a microscopic traffic simulation software, is employed to create a meticulously detailed model of the chosen freeway section. This virtual environment enabled the generation of a rich variety of realistic traffic scenarios, encompassing both normal and abnormal traffic conditions. VISSIM is used to overcome limitations associated with directly simulating incidents within the software [132]. Incidents are generated by scheduling vehicles to make full stops in predetermined locations for a certain duration (20 min in this study). This approach allows for the simulation of various incident severities by strategically blocking a different number of lanes. Additionally, the distance between traffic detectors and the locations of the incidents are meticulously adjusted to reflect a wider range of real-world possibilities. Further, variations in congestion level were modeled by adjusting the ratio of demand to capacity (D/C), with higher ratios signifying increased congestion. Incident severity was manipulated by altering the number of blocked lanes. The distance between traffic detectors was varied by adjusting their locations, and incident locations were diversified to capture broader real-world scenarios. Furthermore, data collection meticulously captures distinct phases—before, during, and after incidents—to effectively capture the critical transition between normal and disrupted traffic flow. This comprehensive approach ensures the dataset accurately reflects the diverse range of incident scenarios encountered in practice. It is important to acknowledge that while simulated data offers advantages such as controlled manipulation of specific variables, validation with real-world data remains crucial. Simulated data may not fully capture the subtleties of real-world driver behavior and environmental factors. Therefore, future validation with real-world data is considered essential for further model refinement.
Traffic information, including speed, volume, and occupancy, is meticulously collected from detectors upstream and downstream of the incident at 30-s intervals. Each simulated scenario lasts an hour and a half, with a 15-min warm-up phase followed by a period of stable data collection. Variations in traffic flow are addressed by running multiple simulations with different seed numbers. This injects randomness into the simulation process, helping to account for the unpredictable nature of real-world traffic conditions. Reliable averages are obtained using a trimmed mean approach to minimize the influence of outliers. The trimmed mean approach removes a small, predetermined percentage of the highest and lowest values from each data set before calculating the average [133,134,135]. This technique helps to mitigate the effects of extreme values that may skew the overall results and provide a more accurate representation of the typical traffic patterns. Normal traffic conditions are simulated by running scenarios without incidents; traffic parameters are collected throughout these scenarios.

3.4. Characteristics of the Generated Dataset

The data generation process resulted in a comprehensive dataset encompassing a wide range of traffic conditions. The dataset consists of 150 individual scenarios, with 22 representing normal traffic flow (without incidents) and the remaining 128 containing simulated incidents.
For each scenario, traffic data—including speed, volume, and occupancy—was meticulously collected at 30-s intervals over a one-hour period. This translates to 120 data intervals per scenario, resulting in a total of 18,000 data intervals across the entire dataset. Within this collection, 12,800 intervals represent normal traffic conditions, while the remaining 5120 intervals correspond to scenarios with simulated incidents.
To ensure robust model evaluation and mitigate overfitting, an 80/20 train-test split was implemented using a 5-fold cross-validation approach. This technique involves dividing the data into five folds [136,137,138,139,140]. Four folds are used for training and validation purposes, while the remaining fold is used for testing. This process is repeated five times, ensuring that each data point is used for testing once. By leveraging this approach, the model’s performance is assessed on unseen data, promoting generalizability and reducing the risk of the model being overly tailored to the training data.

3.5. Development of the AID Model Using Multi-Layer Feedforward Artificial Neural Network (MLFANN)

The proposed AID model leverages a Multi-Layer Feedforward Artificial Neural Network (MLFANN). MLFANN is a specific type of ANN architecture characterized by a layered structure. It consists of an input layer, one or more hidden layers, and an output layer [141,142]. Data flows forward through the network, starting from the input layer, where it is received by the neurons. The neurons in each layer calculate a weighted sum of the inputs they receive, given by the formula:
z j = i = 1 n w i j x i + b j ,
where z j is the net input to neuron j, w i j   represents the weight, x i is the input, and b j is the bias term.
These neurons process the data using activation functions such as the sigmoid function, given by the formula below:
a j = 1 1 + e z j
a j is the result of applying the activation function to z j which is then sent to the next layer for further processing. This process continues until the final output layer is reached, where the processed information is delivered as the model’s prediction. Additionally, during the training phase, an error calculation is performed at the output layer. At the output layer, the model generates a prediction, which, for binary classification, is calculated using the SoftMax function:
y k = e z k j = 1 m e z j
where y k is the predicted probability for class k, based on the weighted sum z k .
This error is then propagated backward through the network, adjusting the weights between neurons in an iterative process using gradient descent:
w i j ( t + 1 ) = w i j ( t ) η L w i j
where η is the learning rate, and L w i j is the gradient of the loss with respect to the weight. These weight adjustments aim to minimize the overall error and optimize the model’s performance.
The selection of MLFANN for this AID model is driven by its success in previous studies. MLFANNs have demonstrated promising results in AID applications, achieving high DR and low FAR and MTTD [87,88,89,90]. Additionally, MLFANNs offer several advantages, including their ability to learn complex non-linear relationships within data, making them well-suited for modeling the intricate dynamics of traffic flow [137,138,139,140,141,142].
RapidMiner, a data mining software platform [143], was employed to develop and fine-tune the MLFANN model. Traffic data collected from upstream and downstream detectors served as the model’s input, and the model was trained to classify the traffic state as either normal or incident.
A crucial aspect of this work involved meticulous fine-tuning of the MLFANN model’s hyperparameters. These hyperparameters are settings that influence the learning process but are not directly learned by the model itself. Examples include the number and size of hidden layers, learning rate, momentum, error tolerance, and training epochs (iterations). Optimizing these hyperparameters plays a vital role in maximizing model effectiveness and addressing potential issues such as underfitting, overfitting, and slow convergence [144,145]. The objective of the fine-tuning process was to maximize the F-score, which represents a harmonic mean of precision and recall, ensuring that the model strikes a balance between DR and FAR. This balance is particularly important in incident detection models, where a high detection rate is necessary but must be achieved while minimizing false alarms. As a result of the fine-tuning process, the final configuration for the MLFANN model was established. The model utilizes a single hidden layer with 35 neurons, optimized to capture non-linear relationships between the 16 input variables. These input variables include traffic flow, speed, and occupancy data from upstream and downstream stations, along with their differences and relative values. The model processes these variables to detect changes in traffic conditions and classify the traffic state as either “incident” or “normal”. The learning rate of 0.015 allows the model to make gradual adjustments during training, avoiding drastic changes that could lead to instability, while the momentum of 0.9 helps accelerate learning and avoid local minima. The error tolerance of 1.00 × 10−10 ensures that training stops only when the error is extremely small, maximizing the model’s accuracy. The model was trained over 1000 epochs, providing sufficient time for the weights to converge and the model to learn effectively from the training data. These choices were made to ensure that the model performs optimally under diverse traffic conditions while minimizing errors. By fine-tuning these hyperparameters, the model is able to balance incident detection accuracy with efficiency, producing reliable predictions across different scenarios. It is important to note that a more in-depth exploration of the hyperparameter optimization process is presented in separate publications by the authors for thoroughness [145].

4. Results

This chapter presents the results obtained from the developed AID model utilizing MLFANN. The performance of the model during both the cross-validation and testing phases is analyzed. Here, the focus is on evaluating the model’s effectiveness in classifying traffic conditions as normal or incident. A sensitivity analysis is then conducted to investigate how each of the four key factors incorporated into the model (traffic congestion level, incident severity, location of the incident, and distance between traffic detectors) impacts the model’s overall performance. Finally, the results of the proposed model are compared with existing AID models documented in the literature. This comparative analysis is undertaken to assess the efficacy of the developed model and its contribution to advancements in the field of AID systems.

4.1. Cross-Validation and Testing Phases Results

To assess the effectiveness of the developed AID model, a 5-fold cross-validation approach was employed. This technique rigorously evaluates model performance by dividing the data into five folds [136,144]. Four folds are used for training and validation, and the remaining fold is used for testing. This process is repeated five times, ensuring each data point is used for testing once. This approach helps mitigate overfitting and promotes model generalizability. A confusion matrix is a valuable tool for visualizing the performance of a classification model. It provides a breakdown of how the model classified the data points, including the number of correctly classified instances (True Positives and True Negatives) and incorrectly classified instances (False Positives and False Negatives) as illustrated in Table 1.
While confusion matrices are a valuable tool to assess model performance by showing the correctly classified and incorrectly classified instances, they can misrepresent real-world performance due to factors such as fluctuating incident alarms and consecutive false alarms. To address this, this study adopts the following assumptions:
Time To Detect (TTD) is considered the first interval at which the model correctly detects an incident. This assumption reflects real-world practices where incident alarms trigger verification procedures, such as camera monitoring, to confirm their authenticity. Therefore, the initial detection of a potential incident is the most crucial aspect.
Consecutive false alarms lasting four or fewer intervals (two minutes or less) are treated as a single false alarm. Traffic operators in real-world scenarios verify consecutive alarm sequences through visual inspections. Short-lived, consecutive false alarms are often disregarded to avoid overwhelming operators and potentially missing critical subsequent incidents. A four-interval threshold (two minutes) balances the need to capture real incidents while mitigating the influence of fleeting false alarms on FAR.
The cross-validation process evaluated the model’s performance on a dataset encompassing 121 scenarios. This dataset included 22 normal traffic conditions and 99 incident scenarios. The normal scenarios consisted of 2640 intervals, while the incident scenarios comprised 11,880 intervals. The model achieved an impressive DR of 94.96% on the dataset of 99 incidents, with only five incidents going undetected. These undetected incidents were all minor lane blockages that occurred during periods of low traffic volume, contributing to the model’s inability to identify them. Analyzed across 121 h of traffic data assessed at 30-s intervals (14,520 model applications), the model generated 147 false alarms, yielding a FAR of 1.01%.
Encouragingly, the model’s performance on the testing dataset indicated improvement in incident detection compared to the cross-validation results. The model successfully identified all incidents in the testing dataset, achieving a 100% DR. It is noteworthy that this dataset included two incidents with a 0.6 D/C ratio and one lane blockage severity, similar to the undetected incidents in the cross-validation set. This suggests that the model learned from its experiences during cross-validation and was able to better classify such incidents in the testing phase. The slight variations in performance between the cross-validation and testing phases are likely attributable to factors such as the specific positioning and spacing of the detectors used in each data collection process.
On the testing dataset, the MTTD increased to 1.6 min. Traffic measurements were collected for 29 h during the testing phase, resulting in 3480 model applications. The model generated 45 false alarms within these intervals, yielding a FAR of 1.29%. This increase in FAR compared to the cross-validation phase aligns with observations in prior literature, where a rise in DR is often accompanied by a corresponding increase in FAR.
A more detailed analysis of these performance metrics and their influencing factors is presented in the following subsections.

4.2. Investigating the Influence of Traffic Congestion Level (D/C Ratio) on Model Performance

In order to gain an understanding of how traffic congestion impacts the model’s performance, the D/C ratio was systematically varied while the other three factors (incident severity, location, and detector spacing) were held constant. This approach isolates the effect of congestion on DR, FAR, and MTTD.
The D/C ratio was set at four distinct values: 0.6, 0.8, 1.0, and 1.2. These values represent a spectrum of traffic congestion levels, ranging from a low demand of 60% capacity (0.6) to a congested scenario exceeding capacity (1.2). By analyzing the model’s performance at each D/C ratio, an assessment can be made of how congestion affects its ability to accurately detect incidents. The model achieved a DR of 100% for congestion levels corresponding to D/C ratios of 0.8, 1.0, and 1.2. However, for the lowest congestion level (D/C ratio of 0.6), the DR dropped to 76.2%. Figure 2 shows the relation between the D/C ratios and DR. This decrease can be attributed to the model’s failure to detect five specific incident cases that occurred during this low-traffic scenario. Notably, all five undetected incidents involved only one lane blockage, representing a minority of incident types. Consequently, while the low traffic volume likely played a role in the model’s difficulty in identifying these minor incidents, the impact of incident severity on detection rates will be further investigated in the next subsection.
As depicted in Figure 2 below, excluding these minor incidents from the analysis, the DR remains 100% for congestion levels across all D/C ratios.
The analysis of FAR revealed a non-monotonic relationship with the D/C ratio. At the lowest congestion level (D/C ratio of 0.6), the FAR was 0.93%. Interestingly, the FAR increased to 1.51% as the congestion level rose to a D/C ratio of 0.8. This peak in FAR can be explained by the operational challenges at near-capacity conditions. With limited space and restricted lane maneuverability, the model might misinterpret minor traffic disruptions, such as lane changes, as incidents, leading to more false alarms. Conversely, when the D/C ratio reaches 1.0 and 1.2, signifying congested scenarios with vehicles moving in platoons, incidents become more distinct and easier for the model to detect. This is reflected in the relatively stable FAR values of 0.944% and 0.94% observed at these higher congestion levels, as shown in Figure 3. It is important to note that including the minor, one-lane blockage incidents at the 0.6 D/C ratio resulted in a slightly lower FAR of 0.873% compared to excluding them. However, this decrease is likely due to comparing the number of false alarms to a larger number of total application intervals when these minor incidents were included. Conversely, excluding these incidents led to a slight increase in FAR (0.938%) as the number of false alarms was compared to a smaller number of application intervals.
The analysis depicted in Figure 4 reveals a positive correlation between D/C ratio (congestion level) and MTTD. As congestion increased from 0.6 to 1.2, the MTTD rose from 0.25 min to around 1 min. This observed increase can be attributed to the formation and presence of vehicle queues at higher congestion levels. Traffic flow becomes more sluggish, with vehicles traveling in platoons. This, in turn, results in longer travel times and delays the propagation of the incident’s impact downstream towards the detectors. Consequently, there is a time lag before the model can detect the incident, leading to a higher MTTD.

4.3. Quantifying the Impact of Incident Severity on Model Performance

This subsection investigates the influence of incident severity on the model’s performance, specifically its impact on DR, FAR, and MTTD. Incident severity is varied by considering lane blockages ranging from one lane (least severe) to five lanes (most severe).
The analysis revealed the most significant influence of incident severity on DR occurred at the one-lane blockage level. In this scenario, the model achieved a DR of approximately 80%. Conversely, for incidents involving three and five lane blockages (representing higher severity), the model maintained a perfect DR of 100%. It is noteworthy that excluding the previously discussed one-lane blockage incidents occurring at a D/C ratio of 0.6 (low traffic volume) results in a 100% DR across all lane blockage severities, including one-lane blockages at higher D/C ratios (0.8, 1.0, and 1.2). This observation reinforces the conclusion from the previous subsection: the combination of low incident severity and low traffic volume makes these incidents challenging for the model to detect. These findings align with previous research in the field, which suggests that minor incidents often have minimal impact on traffic flow and might go undetected [32,52,106,132,146]. Therefore, for the remainder of the analysis, one-lane blockage incidents that occurred at the 0.6 D/C ratio (low traffic volume) will be excluded due to their negligible impact on traffic flow and the model’s performance.
Interestingly, the analysis revealed minimal sensitivity of FAR to variations in incident severity. The FAR remained relatively constant across the three lane blockage scenarios, with values of 1.23%, 1.01%, and 1.21% for one, three, and five lane blockages, respectively. This suggests that the number of lanes blocked by an incident has little influence on the model’s propensity to generate false alarms.
On the other hand, the analysis of MTTD revealed a contrasting trend compared to DR. In this case, MTTD exhibited a decreasing pattern as incident severity increased. For one-lane blockages (least severe), the MTTD was observed to be around 2.26 min. This value steadily decreased to approximately 0.6 min for five-lane blockages (most severe).
This observation can be explained by the growing impact of incident severity on traffic flow. As the number of blocked lanes increases, the incident disrupts traffic flow more significantly, causing greater turbulence and delays. Consequently, the model can detect these more severe incidents faster, resulting in a lower MTTD. This trend aligns with previous research findings, which suggest that incidents with higher severity and a more prominent impact on traffic flow are typically detected quicker by AID systems [87,111].

4.4. Sensitivity Analysis of Model Performance to Detector Spacing

This subsection explores the influence of detector spacing on the model’s performance, focusing on DR, FAR, and MTTD. Three distinct spacings were evaluated: 500 m, 1 km, and 1.5 km.
Similar to the previous analyses, DR exhibited minimal sensitivity to detector spacing. Excluding the previously discussed minor incidents (one-lane blockages at 0.6 D/C ratio), the model achieved a perfect DR of 100% for all incident scenarios and detector spacings. This indicates that the model’s ability to detect incidents remains unaffected by the distance between upstream and downstream detectors.
In contrast to DR, both FAR and MTTD displayed a positive correlation with detector spacing. As the spacing increased from 500 m to 1.5 km, FAR rose from approximately 0.8% to 1.7%. Similarly, MTTD exhibited an upward trend, increasing from 0.5 min to 1.56 min.
This observed trend can be attributed to several factors. With larger spacings between detectors, the traffic characteristics measured upstream and downstream might diverge due to lane changes, merging/diverging traffic, or weaving maneuvers. These variations can be misinterpreted as incidents by the model, leading to a higher number of false alarms. Additionally, the increased travel time between detectors caused by the larger spacing can delay the detection of actual incidents, resulting in a higher MTTD.
Furthermore, the more complex traffic patterns that emerge with longer detector spacings, such as weaving and merging maneuvers, can pose challenges for timely incident detection. Reduced sensor density due to larger gaps between detectors can also contribute to the increase in MTTD. These observations align with previous research findings, such as those reported by c et al. [106], who documented a rise in MTTD with increased detector spacing.
The analysis suggests that smaller detector spacings lead to improved performance in terms of FAR and MTTD. Closer proximity of detectors enables faster incident detection and reduces potential discrepancies in traffic measurements between upstream and downstream locations. However, it is crucial to acknowledge the practical limitations associated with smaller spacings. Installation and maintenance costs can increase significantly with denser detector deployments.
Conversely, larger detector spacings offer a more cost-effective and easily maintainable alternative. However, this comes at the expense of higher FAR and MTTD values, as discussed earlier and documented in previous studies [106,146].
Therefore, the selection of an optimal detector spacing necessitates a careful evaluation of specific application requirements, available resources, and the inherent trade-offs between performance metrics (DR, FAR, and MTTD) and associated costs. Striking a balance between these factors is paramount for developing an incident detection system that is both effective and cost-efficient.

4.5. Evaluating the Effect of Incident Location on Model Performance

This subsection investigates the influence of incident location on the model’s performance in terms of DR, FAR, and MTTD. Nine distinct incident locations were considered, spanning three detector spacings (500 m, 1 km, and 1.5 km), with incidents positioned at quarter (0.25), half, and three-quarter (0.75) distances between the detectors.
The analysis revealed a consistent DR of 100% across all incident locations, regardless of detector spacing. This finding highlights the model’s ability to effectively detect incidents irrespective of their position on the roadway segment monitored by the detectors.
Interestingly, FAR exhibited a decreasing trend as the incident location moved further away from the upstream detector. When incidents occurred closer to the upstream detector (quarter distance), the FAR was observed to be around 1.36%. This value progressively decreased to approximately 0.74% as the incident location shifted towards the downstream detector (three-quarter distance).
This downward trend can be explained by the time it takes for the incident’s impact to propagate upstream. Incidents closer to the upstream detector cause quicker disruptions to traffic flow, which the model might misinterpret as incidents in some cases, leading to false alarms. Conversely, incidents positioned further downstream take longer to affect traffic flow measured by the upstream detector. This delayed impact reduces the likelihood of the model mistaking normal traffic fluctuations for incidents, resulting in fewer false alarms.
The analysis revealed a positive correlation between MTTD and the distance of the incident from the upstream detector. Incidents closer to the upstream detector were detected faster, with an MTTD of approximately 0.85 min. This value gradually increased to around 1.15 min for incidents positioned near the downstream detector. This observation aligns with the explanation for the decreasing FAR. The delayed propagation of the incident’s impact upstream translates to a longer time for the model to detect the incident, hence the higher MTTD for incidents further downstream. This inverse relationship between FAR and incident location further reinforces the concept that the model is less likely to misinterpret normal traffic flow as incidents when the incident’s effect takes longer to reach the upstream detector.

5. Discussion

In the ensuing section, a comprehensive comparison is drawn between the performance of the developed model, focusing on DR, FAR, and MTTD, and those of notable existing models in the literature. Table 2 encapsulates the essence of this comparative analysis.
The developed model demonstrates a well-rounded performance profile when compared to existing AID models, as summarized in Table 2. It achieves a high DR of 95.96%, indicating its effectiveness in identifying incidents. This is coupled with a low FAR of 1.01%, minimizing unnecessary alerts that disrupt traffic flow. The model also boasts an acceptable MTTD of 0.89 min, ensuring timely incident response.
A key strength of the developed model lies in its balanced performance. Unlike some models (e.g., Rossi et al. [106]) that focus on limited parameters, this model considers a wider range of factors influencing traffic flow. This comprehensive approach leads to better generalizability, allowing the model to adapt to various real-world traffic scenarios without relying heavily on specific conditions.
Furthermore, the analysis addresses potential limitations observed in existing models. Certain models, such as the one proposed by Xie et al. [39], achieve high DR and low FAR but lack MTTD values. The use of synthetic incident data for training in these models might lead to overfitting, hindering their performance in real-world situations with greater variation. Additionally, some models (e.g., Zyryanov [5]) only report DR, neglecting FAR and MTTD, making it difficult to comprehensively assess their effectiveness. Video-based models (e.g., Ren et al. [121]) may achieve comparable DR, FAR, and MTTD, but their practicality can be limited by factors such as lighting conditions, extreme weather, and computational demands.
It is important to acknowledge that the developed model’s training and testing relied on simulated traffic data. While the results are promising, future validation with real-world traffic data is recommended for broader applicability. Overall, the developed model offers a competitive advantage with its balanced performance, comprehensiveness, and generalizability, making it a valuable tool for incident detection in real-world traffic management applications.

6. Summary and Conclusions

6.1. Summary

Traffic incidents are a leading cause of fatalities and congestion on roadways worldwide. Since the 1970s, researchers have strived to develop AID models that efficiently and promptly identify incidents. These models play a crucial role in mitigating the negative consequences of incidents by enabling faster response times and improved traffic management strategies. However, existing AID models often suffer from limitations. They may focus on a limited set of factors influencing traffic flow, neglecting the complex interplay between these factors. Additionally, the lack of real-world data encompassing a diverse range of traffic scenarios can hinder the development of truly robust and generalizable models.
This research addresses these limitations by proposing a novel and realistic AID model. This MLFANN model is designed to be comprehensive, considering a wider range of traffic flow parameters simultaneously. These parameters include traffic volume, speed, occupancy, congestion levels, distances between detectors, incident locations, and incident severity. By incorporating these factors, the model offers a more realistic representation of real-world traffic dynamics.
To overcome the scarcity of real-world data with diverse traffic scenarios, VISSIM, traffic simulation software, was employed to generate a comprehensive dataset. These data encompassed various incident scenarios, ensuring the model’s exposure to a wide range of traffic conditions. Additionally, a sensitivity analysis was conducted to isolate and analyze the impact of each individual factor on the model’s performance, measured by DR, FAR, and MTTD.

6.2. Conclusions

The developed MLFANN model exhibited well-rounded performance, achieving a high DR of approximately 96% and a low FAR of around 1%, indicating its effectiveness in accurately identifying incidents while minimizing disruptions caused by false alarms. Furthermore, the model demonstrated an acceptable MTTD of around 0.9 min, facilitating a timely response to incidents. These results compare favorably with existing models that often struggle to achieve such a balanced performance profile.
The sensitivity analysis conducted in this research shed light on the critical factors influencing the model’s performance. This analysis provides valuable insights for future research and real-world deployment.
  • Mitigating Low-Impact Incidents: During periods of low traffic volume, minor incidents can be challenging to detect due to their minimal impact on traffic flow. This aligns with previous research [32,106,131,146,147]. The model relies on significant deviations in traffic patterns to identify incidents, and minor events during low traffic may not cause sufficient disruption to trigger an alarm.
  • The Duality of Congestion: Congestion levels (D/C ratio) exhibit a two-fold effect. While high congestion contributes to a decrease in FAR, it can also lead to longer MTTD values. During peak hours, consistent traffic patterns make it easier for the model to identify abnormal behavior indicative of incidents (lower FAR) [106,108,146].However, queues forming at blocked sections can delay the overall impact on traffic flow, resulting in higher MTTD.
  • Severity’s Impact on Detection Speed: The severity of an incident plays a significant role in detection times. Incidents with more severe lane blockages exert a greater influence on traffic flow, acting as readily detectable signals for the model. This translates to shorter MTTD values, as these incidents are easier to identify [87,109].
  • Distance and Detection Time: The distance between the incident location and the upstream detector significantly impacts detection time. As this distance increases, the incident’s impact takes longer to propagate upstream, leading to higher MTTD values [106,148]. Conversely, incidents further from the detector can experience a decrease in FAR as their delayed impact reduces the likelihood of false detections.
  • Balancing Detector Spacing: Detector spacing necessitates a balancing act. Larger spacings, while potentially offering cost-effectiveness, can contribute to longer MTTD due to delays in incident detection, as observed in previous research [106,148]. Conversely, smaller spacings may lead to an increase in FAR due to fluctuations in traffic measurements caused by longer travel times between detectors.
  • Optimizing Persistence Testing: The research emphasizes the importance of persistence testing to mitigate false alarms. While treating consecutive false alarms as a single event (if they persist for a short time) helps reduce FAR, it is crucial to acknowledge the potential impact on incident detection time. This is particularly relevant if an incident occurs during the ignored period.
These findings highlight the importance of considering all these factors simultaneously for robust incident detection. Interestingly, the analysis also revealed the interplay between factors. For example, the analysis suggests that the interaction of congestion level and incident severity can impact the detection of incidents with lower severity during peak hours. These observations align with previous research on incident detection and traffic flow dynamics.
Despite the promising results demonstrated by the model, several limitations should be acknowledged. The complexity and unpredictability of real-world traffic conditions, including sudden driver actions, oversize vehicles, and varying vehicle speeds, introduce variables that the current model does not fully account for. These factors could impact the system’s accuracy and effectiveness in consistently predicting and preventing incidents. Future work will focus on integrating these variables into the model, along with further testing using real-world data, to improve generalizability and robustness under dynamic traffic conditions.

6.3. Recommendations for Future Research

This research highlights the potential of AI models for improving traffic safety. Here are recommendations to further enhance incident detection models, the developed model in this paper, and overall road safety strategies:
Recommendations for the Developed Model:
  • Real-World Testing: Validate the model’s performance using extensive real-world traffic data to assess its effectiveness in practical settings.
  • Model Generalizability: Evaluate the model’s performance across various freeways, highway systems, and traffic conditions to determine its generalizability.
  • Advanced Persistence Algorithms: Develop and evaluate more sophisticated persistence tests or algorithms to further reduce False Alarm Rates (FAR) and improve the model’s overall reliability.
Enhancing Incident Detection Models. Explore incorporating data from emerging technologies such as:
  • Connected Vehicles: Leverage real-time data from connected vehicles to gain deeper insights into traffic flow, vehicle health, and driver behavior.
  • Advanced Sensors: Utilize advanced sensors such as LiDAR and high-resolution cameras to improve detection accuracy and identify various incident types.
  • Big Data Analytics: Employ big data analytics to analyze vast datasets and uncover hidden patterns that can aid in incident prediction and prevention.
  • Multi-Source Data: Consider incorporating data beyond traditional traffic flow parameters. Explore integrating weather data, road condition reports, and social media feeds to capture a more holistic view of the traffic environment.
  • Transfer Learning: Investigate transfer learning techniques to leverage pre-trained models on related tasks, reducing training time and effort.
  • Explainable AI: Develop models that provide explanations for their decisions. This transparency can enhance trust and facilitate improvements.
Practical Considerations:
  • Cost-Effectiveness: Balance model complexity with cost. Explore cost-effective sensor deployment strategies and efficient computational resources for real-world implementation.
  • Scalability: Design models that are scalable to accommodate diverse road networks and traffic patterns.
  • Real-World Validation: Rigorously test models with real-world traffic data to ensure their effectiveness and generalizability.
  • Potential challenges of model deployment: While the developed model shows promising results, deploying it in real-world traffic management systems may present certain challenges. These include hardware requirements, such as ensuring sufficient computational power for real-time data processing, particularly in systems that rely on edge computing for rapid incident detection. Additionally, reliable data transmission is crucial, especially in regions with limited network infrastructure where sensor data must be consistently transmitted to control centers. Finally, managing processing time is essential for timely incident detection and response, which may require optimizing the model’s complexity to balance accuracy and computational efficiency. Addressing these challenges will be important for the practical implementation of the model in traffic management systems.
General Recommendations for Road Safety:
Proactive Measures:
  • Vehicle Inspections: Implement mandatory and regular vehicle inspections to identify potential mechanical issues before they cause breakdowns or accidents.
  • Road Maintenance: Prioritize regular road inspections and maintenance to address infrastructure deficiencies that contribute to accidents (e.g., potholes, inadequate signage).
  • Driver Education: Promote driver education programs to enhance awareness of traffic safety rules, defensive driving techniques, and the importance of responsible driving behavior.
Public Awareness Campaigns:
  • Severe Weather Alerts: Disseminate timely and clear public alerts through various channels (e.g., media, mobile apps) to warn drivers about severe weather conditions and advise on safe driving practices.
  • Incident Rerouting: Utilize real-time traffic data to provide drivers with dynamic rerouting alerts, minimizing congestion and reducing the likelihood of secondary incidents.
  • Leverage data from connected vehicles, advanced sensors, and big data analytics to gain deeper insights and improve detection accuracy.
  • Consider incorporating multi-source data such as weather, road conditions, and social media feeds for a more holistic view.
  • Explore transfer learning and explainable AI techniques to improve model efficiency and trust.
  • Focus on cost-effective sensor deployment, efficient computational resources, and model scalability for real-world implementation.
  • Validate models with extensive real-world data to ensure their effectiveness and generalizability.
  • Real-World Validation: Extensive testing with real-world traffic data is essential to comprehensively assess the model’s effectiveness and reliability in practical settings.
  • Model Generalizability: Investigating the model’s performance across diverse freeways and highway systems will evaluate its generalizability to different traffic conditions and incident scenarios.
  • Advanced Persistence Algorithms: Developing and evaluating more sophisticated persistence tests or algorithms can further reduce FAR and improve the overall reliability of the incident detection model.
  • Integration with Traffic Management Strategies: Exploring the integration of incident detection models with advanced Intelligent Transportation systems, which can optimize traffic flow and alleviate congestion, leading to improved overall transportation efficiency.
  • Collaboration with Stakeholders: Close collaboration with transportation agencies and stakeholders will ensure the model aligns with operational requirements and can be seamlessly integrated into existing infrastructure.
  • Cost-Benefit Analysis: A comprehensive cost-benefit analysis is crucial to evaluating the economic feasibility of implementing the developed model. This analysis should consider initial investments, operational costs, potential savings from reduced congestion and improved safety, and the overall impact on the transportation network.
  • Emerging Technologies: Leveraging data from autonomous and connected vehicles offers valuable insights and enables more accurate and timely incident detection, leading to proactive traffic management strategies.
  • Minor Incident Reporting Systems: Implementing user-friendly mobile applications or dedicated hotlines for reporting minor incidents during low traffic volume periods will aid in their detection and response.
  • Detector Placement Optimization: Studies to determine the ideal detector spacing that balances detection accuracy and cost-effectiveness are recommended. This optimization will enhance incident detectability and response time, contributing to improved overall traffic management.
Implementing these recommendations would enable the development of more robust and effective incident detection models. Additionally, the performance of the model presented in this paper could be improved. Ultimately, this would lead to the creation of safer and more efficient transportation systems. It is important to note, however, that the most effective approach might involve a combination of incident detection and proactive measures designed to reduce the root causes of accidents.

Author Contributions

Conceptualization, O.E. and A.A.; methodology, O.E. and A.A.; software, O.E.; validation, O.E. and A.A.; formal analysis, O.E.; investigation, O.E.; resources, O.E.; data curation, O.E.; writing—original draft preparation, O.E.; writing—review and editing, A.A.; visualization, O.E.; supervision, A.A.; project administration, A.A.; funding acquisition, A.A. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the American University of Sharjah through a Graduate Teaching Assistantship (GTA) Provided by the Office of Research and Graduate Studies as part of the support to the PhD Program in Engineering Systems Management.

Data Availability Statement

The cross-validation and testing datasets utilized for developing the model in this journal paper are accessible at the following link: https://www.dropbox.com/scl/fo/qbalri06pqpheqbopj0ce/h?rlkey=n9q67igec17iu7b1erq0gl15l&st=g52or3ey&dl=0 (accessed on 20 September 2024).

Acknowledgments

The work in this paper was supported, in part, by the Open Access Pro-gram from the American University of Sharjah. This paper represents the opinions of the author(s) and does not mean to represent the position or opinions of the American University of Sharjah.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Kamran, S.; Haas, O. A Multilevel Traffic Incidents Detection Approach: Identifying Traffic Patterns and Vehicle Behaviours using real-time GPS data. In Proceedings of the 2007 IEEE Intelligent Vehicles Symposium, Istanbul, Turkey, 13–15 June 2007; IEEE: Piscataway, NJ, USA, 2007; pp. 912–917. [Google Scholar]
  2. Srinivasan, D.; Jin, X.; Cheu, R.L. Evaluation of Adaptive Neural Network Models for Freeway Incident Detection. IEEE Trans. Intell. Transp. Syst. 2004, 5, 1–11. [Google Scholar] [CrossRef]
  3. Saini, M. Survey on Vision Based On-Road Vehicle Detection. Int. J. u-and e-Serv. Sci. Technol. 2014, 7, 139–146. [Google Scholar] [CrossRef]
  4. Farradyne, P.B. Traffic Incident Management. In Encyclopedia of Transportation: Social Science and Policy; SAGE Publications, Inc.: Thousand Oaks, CA, USA, 2014. [Google Scholar]
  5. Zyryanov, V.V. Incidents detection on city roads. IOP Conf. Ser. Mater. Sci. Eng. 2020, 913, 042065. [Google Scholar] [CrossRef]
  6. Knoop, V.L.; Hoogendoorn, S.P.; van Zuylen, H.J. Capacity Reduction at Incidents: Empirical Data Collected from a Helicopter. Transp. Res. Rec. 2008, 2071, 19–25. [Google Scholar] [CrossRef]
  7. World Health Organization. [Internet]. 2021 [Cited 2021 Dec 21]. Road Traffic Injuries. Available online: https://www.who.int/news-room/fact-sheets/detail/road-traffic-injuries (accessed on 21 December 2023).
  8. Independent Evaluation Group (IEG). Making Roads Safer: Learning from the World Bank’s Experience; IEG Learning Note; World Bank: Washington, DC, USA, 2014. [Google Scholar]
  9. Yuan, F.; Cheu, R.L. Incident detection using support vector machines. Transp. Res. Part C Emerg. Technol. 2003, 11, 309–328. [Google Scholar] [CrossRef]
  10. Sheikh, M.S.; Liang, J.; Wang, W. An Improved Automatic Traffic Incident Detection Technique Using a Vehicle to Infrastructure Communication. J. Adv. Transp. 2020, 2020, 9139074. [Google Scholar] [CrossRef]
  11. Li, R.; Pereira, F.C.; Ben-Akiva, M.E. Overview of traffic incident duration analysis and prediction. Eur. Transp. Res. Rev. 2018, 10, 22. [Google Scholar] [CrossRef]
  12. Chimba, D.; Kutela, B.; Ogletree, G.; Horne, F.; Tugwell, M. Impact of Abandoned and Disabled Vehicles on Freeway Incident Duration. J. Transp. Eng. 2014, 140, 04013013. [Google Scholar] [CrossRef]
  13. Valenti, G.; Lelli, M.; Cucina, D. A comparative study of models for the incident duration prediction. Eur. Transp. Res. Rev. 2010, 2, 103–111. [Google Scholar] [CrossRef]
  14. Zhang, H.; Khattak, A. What Is the Role of Multiple Secondary Incidents in Traffic Operations? J. Transp. Eng. 2010, 136, 986–997. [Google Scholar] [CrossRef]
  15. Tantillo, M.J.; Roberts, E.; Mangar, U. Roles of Transportation Management Centers in Incident Management on Managed Lanes; Report No.: FHWA-HOP-14-022; Federal Highway Administration: Washington, DC, USA, 2014.
  16. Jin, X.; Zhang, Z.; Gan, A. Traffic Management Centers: Challenges, Best Practices, and Future Plans; National Center for Transportation Systems Productivity and Management (US): Atlanta, GA, USA, 2014. [Google Scholar]
  17. Xiao, J.; Liu, Y. Traffic Incident Detection Using Multiple-Kernel Support Vector Machine. Transp. Res. Rec. J. Transp. Res. Board 2012, 2324, 44–52. [Google Scholar] [CrossRef]
  18. Allen, R.C.; Cleveland, D.E. The Detection of Freeway Capacity Reducing Incidents by Traffic Stream Measurements; Report No.: TrS-1; Highway Safety Research Institute, The University of Michigan: Ann Arbor, MI, USA, 1970. [Google Scholar]
  19. Dudek, C.L.; Messer, C.J.; Nuckles, N.B. Incident detection on urban freeways. Transp. Res. Rec. 1974, 495, 12–24. [Google Scholar]
  20. Dudek, C.L.; Weaver, G.D.; Ritch, G.P.; Messer, C.J. Detecting Freeway Incidents under Low-Volume Conditions; Transportation Research Record; A & M University: College Station, TX, USA, 1975; Volume 553. [Google Scholar]
  21. Payne, H. Freeway incident detection based upon pattern classification. In Proceedings of the 1975 IEEE Conference on Decision and Control Including the 14th Symposium on Adaptive Processes, IEEE, Houston, TX, USA, 10–12 December 1975; pp. 688–692. [Google Scholar]
  22. Payne, H.J.; Tignor, S. Freeway Incident-Detection Algorithms Based on Decision Trees with States. Transp. Res. Rec. 1978, 682, 30–37. [Google Scholar]
  23. Levin, M.; Krause, G.M. Incident detection: A Bayesian approach. Transp. Res. Rec. 1978, 682, 52–58. [Google Scholar]
  24. Iqbal, Z.; Khan, M.I. Automatic incident detection in smart city using multiple traffic flow parameters via V2X communication. Int. J. Distrib. Sens. Netw. 2018, 14, 1550147718815845. [Google Scholar] [CrossRef]
  25. Parkany, E.; Xie, C. A Complete Review of Incident Detection Algorithms & Their Deployment: What Works and What Doesn’t; Report No. NETCR37, Project No. 00-7; New England Transportation Consortium: Drive Concord, NH, USA, 2005. [Google Scholar]
  26. Fangming, T.; Han, D. Simulation of traffic incident detection based on VISSIM and neural network. In Proceedings of the 2012 IEEE International Conference on Computer Science and Automation Engineering (CSAE), Zhangjiajie, China, 25–27 May 2012; pp. 51–55. [Google Scholar]
  27. Calderoni, L.; Maio, D.; Rovis, S. Deploying a network of smart cameras for traffic monitoring on a “city kernel”. Expert Syst. Appl. 2014, 41, 502–507. [Google Scholar] [CrossRef]
  28. Cheng, H.Y.; Gau, V.; Huang, C.W.; Hwang, J.N. Advanced formation and delivery of traffic information in intelligent transportation systems. Expert. Syst. Appl. 2012, 39, 8356–8368. [Google Scholar] [CrossRef]
  29. Wen, W. An intelligent traffic management expert system with RFID technology. Expert. Syst. Appl. 2010, 37, 3024–3035. [Google Scholar] [CrossRef]
  30. D’Andrea, E.; Marcelloni, F. Detection of traffic congestion and incidents from GPS trace analysis. Expert. Syst. Appl. 2017, 73, 43–56. [Google Scholar] [CrossRef]
  31. Houbraken, M.; Logghe, S.; Schreuder, M.; Audenaert, P.; Colle, D.; Pickavet, M. Automated Incident Detection Using Real-Time Floating Car Data. J. Adv. Transp. 2017, 2017, 8241545. [Google Scholar] [CrossRef]
  32. Liang, Z.; Chen, H.; Song, Z.; Zhou, Y.; Zhang, B. Traffic congestion incident detection and dissipation algorithm for urban intersection based on FCD. In Proceedings of the 2017 3rd IEEE International Conference on Computer and Communications (ICCC), Chengdu, China, 13–16 December 2017; pp. 2578–2583. [Google Scholar]
  33. Asakura, Y.; Kusakabe, T.; Nguyen, L.X.; Ushiki, T. Incident detection methods using probe vehicles with on-board GPS equipment. Transp. Res. Part C Emerg. Technol. 2017, 81, 330–341. [Google Scholar] [CrossRef]
  34. Ki, Y.K.; Kim, J.H.; Kim, T.K.; Heo, N.W.; Choi, J.W.; Jeong, J.H. Method for Automatic Detection of Traffic Incidents Using Neural Networks and Traffic Data. In Proceedings of the 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), Vancouver, BC, Canada, 1–3 November 2018; pp. 184–188. [Google Scholar]
  35. Roess, R.P.; Prassas, E.S.; McShane, W.R. Traffic Engineering. In ATM: The Broadband Telecommunications Solution, 5th ed.; Institution of Engineering and Technology: London, UK, 1993; pp. 132–147. [Google Scholar]
  36. Elefteriadou, L. An Introduction to Traffic Flow Theory, 17th ed.; Springer Optimization and Its Applications; Springer: New York, NY, USA, 2014; Volume 84, pp. 78–82. [Google Scholar]
  37. Machemehl, R.M. Real Time Freeway Incident Detection; Report No. SWUTC/14/600451-00083-1; Centre for Transportation Research, University of Texas at Austin: Austin, TX, USA, 2014. [Google Scholar]
  38. Motamed, M. Developing a Real-Time Freeway Incident Detection Model Using Machine Learning Techniques. Doctoral Dissertation, University of Texas at Austin, Austin, TX, USA, 2016. [Google Scholar]
  39. Xie, T.; Shang, Q.; Yu, Y. Automated Traffic Incident Detection: Coping with Imbalanced and Small Datasets. IEEE Access 2022, 10, 35521–35540. [Google Scholar] [CrossRef]
  40. Mahmassani, H.S. Evaluation of Incident Detection Methodologies (FHWA/TX-00/1795-1); Report No.: FHWA-OP-99-032; Federal Highway Administration: Austin, TX, USA, 1999.
  41. Martin, P.T.; Perrin, J.; Hansen, B.; Kump, R.; Moore, D. Incident Detection Algorithm Evaluation; Report No.: MPC 01-122; Utah Department of Transportation, Minnesota Department of Transportation: Salt Lake City, UT, USA; Minnesota Department of Transportation: St. Paul, Minnesota, USA, 2001. [Google Scholar]
  42. ElSahly, O.; Abdelfatah, A. A Systematic Review of Traffic Incident Detection Algorithms. Sustainability 2022, 14, 14859. [Google Scholar] [CrossRef]
  43. Levin, M.; Krause, G.M. Incident-Detection Algorithms Part 1. Off-Line Evaluation. Transp. Res. Rec. 1979, 722, 49–58. [Google Scholar]
  44. Levin, M.; Krause, G.M.; Budrick, J.A. Incident-detection algorithms. Part 2. On-line evaluation. Transp. Res. Rec. 1979, 722, 49–58. [Google Scholar]
  45. Liu, Q.; Chung, E.; Zhai, L. Fusing moving average model and stationary wavelet decomposition for automatic incident detection: Case study of Tokyo Expressway. J. Traffic Transp. Eng. (Engl. Ed.) 2014, 1, 404–414. [Google Scholar] [CrossRef]
  46. Bakioğlu, G.; Silgu, M.A.; Özcanan, S.; Gökaşar, I.; Büyük, M.; Çelikoğlu, H.B.; Osman, A. Incident Detection Algorithms: A Literature Review. In Proceedings of the 1st IRF Europe & Central Asia Regional Congress & Exhibition, Istanbul, Turkey, 15–18 September 2015. [Google Scholar]
  47. Lyall, B.B. Performance Evaluation of the McMaster Incident Detection Algorithm; Submitted to the Department of Geography in Fulfillment of the Requirements of Geography 4C06; McMaster University: Hamilton, ON, Canada, 1991. [Google Scholar]
  48. Cohen, S.; Ketselidou, Z. A Calibration Process for Automatic Incident Detection Algorithms. In Proceedings of the 4th International Conference on Microcomputers in Transportation, Baltimore, MD, USA, 22–24 July 1993. [Google Scholar]
  49. Abdulhai, B.; Abdelwahab, H.T. Comparison of three incident detection algorithms using detailed simulation results. J. Transp. Eng. 2001, 127, 251–259. [Google Scholar]
  50. Stephanedes, Y.J.; Hourdakis, J. Comparison of real-time traffic incident detection algorithms. Transp. Res. Rec. 1996, 1554, 44–51. [Google Scholar] [CrossRef]
  51. Collins, J.F.; Hopkins, C.M.; Martin, J.A. Automatic Incident Detection: TRRL Algorithms HIOCC and PATREG; Transport and Road Research Laboratory: Crowthorne, UK, 1979. [Google Scholar]
  52. Masters, P.H.; Lam, J.K.; Wong, K. Incident Detection Algorithms for COMPASS—An Advanced Traffic Management System. In Proceedings of the Vehicle Navigation and Information Systems Conference, Troy, MI, USA, 20–23 October 1991. [Google Scholar]
  53. Balke, K.N. An Evaluation of Existing Incident Detection Algorithms; Report No.: FHWA/TX-93/1232-20; Texas Transportation Institute, Texas A&M University System: College Station, TX, USA, 1993. [Google Scholar]
  54. Deniz, O.; Celikoglu, H.B. Overview to some existing incident detection algorithms: A comparative evaluation. Procedia-Soc. Behav. Sci. 2011, 2, 153–168. [Google Scholar]
  55. Ahmed, M.S.; Cook, A.R. Analysis of Freeway Traffic Time-Series Data by Using Box-Jenkins Techniques. Transportation Research Record 1979. Available online: https://api.semanticscholar.org/CorpusID:106553179 (accessed on 10 September 2024).
  56. Ahuja, L. Automatic Incident Detection; Iowa State University: Ames, IA, USA, 2018. [Google Scholar]
  57. Tsai, J.; Case, E.R. Development of freeway incident-detection algorithms by using pattern-recognition techniques. Transp. Res. Rec. 1979, 722, 113–116. [Google Scholar]
  58. Ahmed, S.A.; Cook, A. Application of Time-Series Analysis Techniques to Freeway Incident Detection. Transp. Res. Rec. 1982, 841, 19–21. [Google Scholar]
  59. Ahmed, S.A.; Cook, A.R. Time Series Models for Freeway Incident Detection. Transp. Eng. J. ASCE 1980, 106, 731–745. [Google Scholar] [CrossRef]
  60. Chakraborty, P.; Hegde, C.; Sharma, A. Data-driven parallelizable traffic incident detection using spatio-temporally denoised robust thresholds. Transp. Res. Part C Emerg. Technol. 2019, 105, 81–99. [Google Scholar] [CrossRef]
  61. Jin, X.; Srinivasan, D.; Cheu, R.L. Classification of freeway traffic patterns for incident detection using constructive probabilistic neural networks. IEEE Trans. Neural Netw. 2001, 12, 1173–1187. [Google Scholar] [CrossRef]
  62. Olugbade, S.; Ojo, S.; Imoize, A.L.; Isabona, J.; Alaba, M.O. A Review of Artificial Intelligence and Machine Learning for Incident Detectors in Road Transport Systems. Math. Comput. Appl. 2022, 27, 77. [Google Scholar] [CrossRef]
  63. Sharma, S.; Harit, S.; Kaur, J. Traffic Accident Detection Using Machine Learning Algorithms. In Proceedings of the Third International Conference on Sustainable Computing. Bosnia and Herzegovina; Springer: Singapore, 2022; pp. 501–507. [Google Scholar]
  64. Rusyaidi, M.; Ibrahim, Z. A Review: An Evaluation of Current Artificial Intelligent Methods in Traffic Flow Prediction. IOP Conf. Ser. Mater. Sci. Eng. 2020, 917, 012063. [Google Scholar] [CrossRef]
  65. Olayode, O.I.; Tartibu, L.K.; Okwu, M.O. Application of Artificial Intelligence in Traffic Control System of Non-autonomous Vehicles at Signalized Road Intersection. Procedia CIRP 2020, 91, 194–200. [Google Scholar] [CrossRef]
  66. Gamel, S.A.; Saleh, A.I.; Ali, H.A. Machine learning-based traffic management techniques for intelligent transportation system: Review. Nile J. Commun. Comput. Sci. 2021, 1, 9–18. [Google Scholar]
  67. Nama, M.; Nath, A.; Bechra, N.; Bhatia, J.; Tanwar, S.; Chaturvedi, M.; Sadoun, B. Machine learning-based traffic scheduling techniques for intelligent transportation system: Opportunities and challenges. Int. J. Commun. Syst. 2021, 34, e4814. [Google Scholar] [CrossRef]
  68. Šusteková, D.; Knutelská, M. How is the artificial intelligence used in applications for traffic management. Mach. Technol. Mater. 2015, 9, 49–52. [Google Scholar]
  69. Yuan, T.; Rocha Neto, W.; Rothenberg, C.E.; Obraczka, K.; Barakat, C.; Turletti, T. Machine learning for next-generation intelligent transportation systems: A survey. Trans. Emerg. Telecommun. Technol. 2022, 33, e4427. [Google Scholar] [CrossRef]
  70. Hamad, K.; Khalil, M.A.; Alozi, A.R. Predicting Freeway Incident Duration Using Machine Learning. Int. J. Intell. Transp. Syst. Res. 2020, 18, 367–380. [Google Scholar] [CrossRef]
  71. Almukhalfi, H.; Noor, A.; Noor, T.H. Traffic management approaches using machine learning and deep learning techniques: A survey. Eng. Appl. Artif. Intell. 2024, 133, 108147. [Google Scholar] [CrossRef]
  72. Suthaharan, S. Machine Learning Models and Algorithms for Big Data Classification; Integrated Series in Information Systems; Springer: Boston, MA, USA, 2016; Volume 36. [Google Scholar]
  73. Mani, D.; Amrith, P.; Umamaheswari, E.; Ajay, D.M.; Anitha, R.U. Smart detection of vehicle accidents using object identification sensors with artificial intelligent systems. Int. J. Recent Technol. Eng. 2019, 7, 375–379. [Google Scholar]
  74. Huang, T.; Wang, S.; Sharma, A. Highway crash detection and risk estimation using deep learning. Accid. Anal. Prev. 2020, 135, 105392. [Google Scholar] [CrossRef] [PubMed]
  75. Gkioka, G.; Dominguez, M.; Tympakianaki, A.; Mentzas, G. AI-Driven Real-Time Incident Detection for Intelligent Transportation Systems. In Advances in Transdisciplinary Engineering; Springer: Berlin, Germany, 2024; pp. 56–68. [Google Scholar]
  76. Usama, M. Application of Machine Learning Techniques for Traffic State Estimation, Pattern Recognition, and Crash Detection; The University of Alabama in Huntsville: Huntsville, AL, USA, 2023. [Google Scholar]
  77. Yijing, H.; Wei, W.; He, Y.; Qihong, W.; Kaiming, X. Intelligent algorithms for incident detection and management in smart transportation systems. Comput. Electr. Eng. 2023, 110, 108839. [Google Scholar] [CrossRef]
  78. Qu, Q.; Shen, Y.; Yang, M.; Zhang, R.; Zhang, H. Expressway Traffic Incident Detection Using a Deep Learning Approach Based on Spatiotemporal Features with Multilevel Fusion. J. Transp. Eng. Part A Syst. 2024, 150, 4024020. [Google Scholar] [CrossRef]
  79. Hopfield, J.J. Neural networks and physical systems with emergent collective computational abilities. Proc. Natl. Acad. Sci. USA 1982, 79, 2554–2558. [Google Scholar] [CrossRef]
  80. Wang, L.; Zhao, J.; Mortier, R. Neural Network. In OCaml Scientific Computing. Undergraduate Topics in Computer Science; Springer: Cham, Switzerland, 2022. [Google Scholar] [CrossRef]
  81. IBM. [Internet]. 2020 [Cited 2021 Dec 27]. What are Neural Networks? Available online: https://www.ibm.com/cloud/learn/neural-networks (accessed on 25 December 2023).
  82. McCulloch, W.S.; Pitts, W. A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. 1943, 5, 115–133. [Google Scholar] [CrossRef]
  83. Hardesty, L.; MIT News|Massachusetts Institute of Technology. [Cited 2021 Dec 27]. p. 2017. Explained: Neural Networks. Available online: https://news.mit.edu/2017/explained-neural-networks-deep-learning-0414 (accessed on 25 December 2023).
  84. Lin, Y.; Li, L.; Jing, H.; Ran, B.; Sun, D. Automated traffic incident detection with a smaller dataset based on generative adversarial networks. Accid. Anal. Prev. 2020, 144, 105628. [Google Scholar] [CrossRef]
  85. Philip, A.O.; Saravanaguru, R.K. Multisource traffic incident reporting and evidence management in Internet of Vehicles using machine learning and blockchain. Eng. Appl. Artif. Intell. 2023, 117, 105630. [Google Scholar] [CrossRef]
  86. Katsamenis, I.; Karolou, E.E.; Davradou, A.; Protopapadakis, E.; Doulamis, A.; Doulamis, N.; Kalogeras, D. TraCon: A Novel Dataset for Real-Time Traffic Cones Detection Using Deep Learning. In Novel & Intelligent Digital Systems: Proceedings of the 2nd International Conference (NiDS 2022); Krouska, A., Troussas, C., Caro, J., Eds.; Lecture Notes in Networks and Systems; Springer International Publishing: Athens, Greece, 2023; Volume 556, pp. 382–391. [Google Scholar]
  87. Cheu, R.L.; Ritchie, S.G. Automated detection of lane-blocking freeway incidents using artificial neural networks. Transp. Res. Part C 1995, 3, 371–388. [Google Scholar] [CrossRef]
  88. Jin, X.; Cheu, R.L.; Srinivasan, D. Development and adaptation of constructive probabilistic neural network in freeway incident detection. Transp. Res. Part C Emerg. Technol. 2002, 10, 121–147. [Google Scholar] [CrossRef]
  89. Dia, H.; Rose, G. Development and evaluation of neural network freeway incident detection models using field data. Transp. Res. Part C Emerg. Technol. 1997, 5, 313–331. [Google Scholar] [CrossRef]
  90. Abdulhai, B.; Ritchie, S.G. Enhancing the universality and transferability of freeway incident detection using a Bayesian-based neural network. Transp. Res. Part C Emerg. Technol. 1999, 7, 261–280. [Google Scholar] [CrossRef]
  91. Cheu, R.L.; Ritchie, S.G.; Recker, W.W.; Bavarian, B. Investigation of a Neural Network Model for Freeway Incident Detection. In Proceedings of the International Conference on the Application of Artificial Intelligence Techniques to Civil and Structural Engineering; University of California, Irvine, Institute of Transportation Studies: Oxford, UK, 1991; pp. 267–274. [Google Scholar]
  92. Gupta, G.; Singh, R.; Singh Patel, A.; Ojha, M. Accident Detection Using Time-Distributed Model in Videos. In Proceedings of the Fifth International Congress on Information and Communication; Yang, X.S., Sherratt, S., Dey, N., Joshi, A., Eds.; Springer: Singapore, 2021; pp. 214–223. [Google Scholar]
  93. Li, L.; Lin, Y.; Du, B.; Yang, F.; Ran, B. Real-time traffic incident detection based on a hybrid deep learning model. Transp. A Transp. Sci. 2022, 18, 78–98. [Google Scholar] [CrossRef]
  94. Ho, T.K. Random decision forests. In Proceedings of the 3rd International Conference on Document Analysis and Recognition, Montreal, QC, Canada, 14–16 August 1995; pp. 278–282. [Google Scholar]
  95. Ruppert, D. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. J. Am. Stat. Assoc. 2004, 99, 567. [Google Scholar] [CrossRef]
  96. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  97. Yiu, T. Understanding Random Forest [Internet]. 2019 [Cited 2022 Aug 3]. Available online: https://towardsdatascience.com/understanding-random-forest-58381e0602d2 (accessed on 12 March 2023).
  98. Random Forest Algorithm [Internet]. 2022 [Cited 2021 May 8]. Available online: https://www.simplilearn.com/tutorials/machine-learning-tutorial/random-forest-algorithm?tag=randomforest (accessed on 12 March 2023).
  99. Dogru, N.; Subasi, A. Traffic accident detection using random forest classifier. In Proceedings of the 2018 15th Learning and Technology Conference (L&T), IEEE, Jeddah, Saudi Arabia, 25–26 February 2018; pp. 40–45. [Google Scholar]
  100. ElSahly, O.; Abdelfatah, A. An Incident Detection Model Using Random Forest Classifier. Smart Cities 2023, 6, 1786–1813. [Google Scholar] [CrossRef]
  101. Zadeh, L.A. Fuzzy sets. Inf. Control 1965, 8, 338–353. [Google Scholar] [CrossRef]
  102. Zadeh, L.A. Fuzzy logic = computing with words. IEEE Trans. Fuzzy Syst. 1996, 4, 103–111. [Google Scholar] [CrossRef]
  103. Zadeh, L.A. Fuzzy algorithms. Inf. Control 1968, 12, 94–102. [Google Scholar] [CrossRef]
  104. Zadeh, L.A. Fuzzy Sets and Systems. Int. J. Gen. Syst. 1990, 17, 129–138. [Google Scholar] [CrossRef]
  105. Nikolaev, A.B.; Sapego, Y.S.; Ivakhnenko, A.M.; Mel’nikova, T.E.; Stroganov, V.Y. Analysis of the incident detection technologies and algorithms in intelligent transport systems. Int. J. Appl. Eng. Res. 2017, 12, 4765–4774. [Google Scholar]
  106. Rossi, R.; Gastaldi, M.; Gecchele, G.; Barbaro, V. Fuzzy Logic-based Incident Detection System using Loop Detectors Data. Transp. Res. Procedia 2015, 10, 266–275. [Google Scholar] [CrossRef]
  107. Ahmed, F.; Hawas, Y.E. A fuzzy logic model for real-time incident detection in urban road network. In Proceedings of the 5th International Conference on Agents and Artificial Intelligence—Volume 2: ICAART, Barcelona, Spain, 15–18 February 2013; pp. 465–472. [Google Scholar]
  108. Mustafa, F.W.F. An Application of Fuzzy Logic in Urban Traffic Incident Detection. Master’s Thesis, United Arab Emirates University, Al Ain, United Arab Emirates, 2015. [Google Scholar]
  109. Lee, S.; Krammes, R.A.; Yen, J. Fuzzy-logic-based incident detection for signalized diamond interchanges. Transp. Res. Part C Emerg. Technol. 1998, 6, 359–377. [Google Scholar] [CrossRef]
  110. Hsu, C.W.; Chang, C.C.; Lin, C.J. A Practical Guide to Support Vector Classification; University of National Taiwan: Taipei, China, 2003. [Google Scholar]
  111. Suthaharan, S. Support Vector Machine. In Machine Learning Models and Algorithms for Big Data Classification; Suthaharan, S., Ed.; Springer: Boston, MA, USA, 2016; pp. 207–235. [Google Scholar]
  112. Kumar, B.; Basit, A.; Kiruba, M.B.; Giridharan, R.; Keerthana, S.M. Road Accident Detection Using Machine Learning. In Proceedings of the 2021 International Conference on System, Computation, Automation and Networking (ICSCAN), Puducherry, India, 30–31 July 2021; pp. 1–5. [Google Scholar]
  113. Ma, Y.; Chowdhury, M.; Sadek, A.; Jeihani, M. Real-Time Highway Traffic Condition Assessment Framework Using Vehicle–Infrastructure Integration (VII) With Artificial Intelligence (AI). IEEE Trans. Intell. Transp. Syst. 2009, 10, 615–627. [Google Scholar]
  114. Xu, M.; Liu, H.; Yang, H. Ensemble learning based approach for traffic incident detection and multi-category classification. Eng. Appl. Artif. Intell. 2024, 132, 107933. [Google Scholar] [CrossRef]
  115. Mahmassani, H.S.; Haas, C.; Zhou, S.; Peterman, J. Evaluation of Incident Detection Methodologies. Doctoral Dissertation, University of Texas at Austin, Austin, TX, USA, 1999. [Google Scholar]
  116. Bartolomé-Hornillos, C.; San-José-Revuelta, L.M.; Aguiar-Pérez, J.M.; García-Serrada, C.; Vara-Pazos, E.; Casaseca-de-la-Higuera, P. A Self-Adaptive Automatic Incident Detection System for Road Surveillance Based on Deep Learning. Sensors 2024, 24, 1822. [Google Scholar] [CrossRef]
  117. Saho, K. Kalman Filter for Moving Object Tracking: Performance Analysis and Filter Design. In Kalman Filters—Theory for Advanced Applications; InTech: Rijeka, Croatia, 2018. [Google Scholar]
  118. Ekstrand, B. Some Aspects on Filter Design for Target Tracking. J. Control Sci. Eng. 2012, 2012, 870890. [Google Scholar] [CrossRef]
  119. Saho, K.; Masugi, M. Automatic Parameter Setting Method for an Accurate Kalman Filter Tracker Using an Analytical Steady-State Performance Index. IEEE Access 2015, 3, 1919–1930. [Google Scholar] [CrossRef]
  120. Hashlamon, I.; Erbatur, K. An improved real-time adaptive Kalman filter with recursive noise covariance updating rules. Turk. J. Electr. Eng. Comput. Sci. 2016, 24, 524–540. [Google Scholar] [CrossRef]
  121. Ren, J.; Chen, Y.; Xin, L.; Shi, J.; Li, B.; Liu, Y. Detecting and positioning of traffic incidents via video-based analysis of traffic states in a road segment. IET Intell. Transp. Syst. 2016, 10, 428–437. [Google Scholar] [CrossRef]
  122. Bao, L.; Wang, Q.; Qu, W.; Mo, X. Research on Highway Traffic Event Detection Method Based on Image Processing. IOP Conf. Ser. Earth Environ. Sci. 2021, 791, 012193. [Google Scholar] [CrossRef]
  123. Fernández, A.; García, S.; Galar, M.; Prati, R.C.; Krawczyk, B.; Herrera, F. Learning from Imbalanced Data Sets; Springer International Publishing: Cham, Switzerland, 2018. [Google Scholar]
  124. Ma, Y.; He, H. Imbalanced Learning; He, H., Ma, Y., Eds.; Wiley: Hoboken, NJ, USA, 2013. [Google Scholar]
  125. Chen, S.; Wang, W.; van Zuylen, H. Construct support vector machine ensemble to detect traffic incident. Expert Syst. Appl. 2009, 36, 10976–10986. [Google Scholar] [CrossRef]
  126. Hamad, K.; Quiroga, C. Geovisualization of Archived ITS Data-Case Studies. IEEE Trans. Intell. Transp. Syst. 2016, 17, 104–112. [Google Scholar] [CrossRef]
  127. Cheu, R.L. Neural Network Models for Automated Detection of Lane-Blocking Incidents on Freeways. In Proceedings of the International Conference on Advanced Technologies in Transportation and Traffic Management, Singapore, 18–20 May 1994; pp. 245–252. [Google Scholar]
  128. Chakraborty, P.; Sharma, A.; Knickerbocker, S.; Hess, J.R.; Sharma, A.; Knickerbocker, S. Outlier Mining Based Traffic Incident Detection Using Big Data Analytics. In Proceedings of the 96th Annual Meeting Transportation Research Board, Washington DC, USA, 8–12 January 2017; pp. 8–12. [Google Scholar]
  129. Ozbay, K.; Kachroo, P. Incident Management in Intelligent Transportation Systems; Artech House Publishers: Boston, MA, USA, 1999; 248p. [Google Scholar]
  130. Karatsoli, M.; Margreiter, M.; Spangler, M. Bluetooth-based travel times for automatic incident detection—A systematic description of the characteristics for traffic management purposes. Transp. Res. Procedia 2017, 24, 204–211. [Google Scholar] [CrossRef]
  131. Highway Capacity Manual, 7th ed.; National Academies Press: Washington, DC, USA, 2022.
  132. Min, S.L. Evaluation of Adaptive Automatic Freeway Incident Detection Algorithms; Malaysia University of Science and Technology: Selangor, Malaysia, 2004. [Google Scholar]
  133. PTV Group. PTV Vissim 2022 User Manual; PTV Group: Karlsruhe, Germany, 2022. [Google Scholar]
  134. Spiegelman, C.H.; Park, E.S.; Rilett, L.R. Transportation Statistics and Microsimulation; CRC Press: Boca Raton, FL, USA, 2011; pp. 33–34. [Google Scholar]
  135. Ngan, V.; Sayed, T.; Abdelfatah, A. Impacts of Various Parameters on Transit Signal Priority Effectiveness. J. Public Transp. 2004, 7, 71–93. [Google Scholar] [CrossRef]
  136. Baturynska, I.; Martinsen, K. Prediction of geometry deviations in additive manufactured parts: Comparison of linear regression with machine learning algorithms. J. Intell. Manuf. 2021, 32, 179–200. [Google Scholar] [CrossRef]
  137. Samarasinghe, S. Neural Networks for Applied Sciences and Engineering; Auerbach Publications: Boca Raton, FL, USA, 2006. [Google Scholar]
  138. Shanmuganathan, S. Artificial Neural Network Modelling; Shanmuganathan, S., Samarasinghe, S., Eds.; Studies in Computational Intelligence; Springer International Publishing: Cham, Switzerland, 2016; Volume 628. [Google Scholar]
  139. Ripley, B.D. Pattern Recognition and Neural Networks; Cambridge University Press: Cambridge, UK, 1996. [Google Scholar]
  140. James, G.; Gareth, M.; Witten, D.; Hastie, T.; Tibshirani, R. An Introduction to Statistical Learning: With Applications in R; Springer: New York, NY, USA, 2013; 426p. [Google Scholar]
  141. Haykin, S.S. Neural Networks: A Comprehensive Foundation, 2nd ed.; Prentice Hall: Upper Saddle River, NJ, USA, 1994. [Google Scholar]
  142. Sazli, M.H. A brief review of feed-forward neural networks. Commun. Fac. Sci. Univ. Ank. 2006, 50, 11–17. [Google Scholar] [CrossRef]
  143. RapidMiner. Available online: https://rapidminer.com/ (accessed on 20 September 2023).
  144. Agrawal, T. Hyperparameter Optimization in Machine Learning, 1st ed.; Apress: Berkeley, CA, USA, 2021. [Google Scholar] [CrossRef]
  145. ElSahly, O.; Abdelfatah, A. Optimizing Hyperparameters of Artificial Neural Network Model for Traffic Incident Detection. In Proceedings of the 50th International Conference on Computers and Industrial Engineering (CIE 50), Sharjah-Dubai, United Arab Emirates, 30 October–2 November 2023. [Google Scholar]
  146. Margreiter, M.; Spangler, M.; Zeh, T.; Carstensen, C. Bluetooth-Measured Travel Times for Dynamic Re-Routing. In Proceedings of the Annual International Conference on Architecture and Civil Engineering (ACE 2015), Singapore, 13–14 April 2015. [Google Scholar]
  147. Ahmed, F.; Hawas, Y.E. A Threshold-Based Real-Time Incident Detection System for Urban Traffic Networks. Procedia-Soc. Behav. Sci. 2012, 48, 1713–1722. [Google Scholar] [CrossRef]
  148. Raosaheb Patil, V.; Suresh Pardeshi, S. Mechanism for accident detection, prevention and reporting system. Mater. Today Proc. 2023, 72, 1975–1980. [Google Scholar] [CrossRef]
Figure 1. Study Area.
Figure 1. Study Area.
Infrastructures 09 00170 g001
Figure 2. Impact of D/C Ratio on DR Excluding Minor Incidents.
Figure 2. Impact of D/C Ratio on DR Excluding Minor Incidents.
Infrastructures 09 00170 g002
Figure 3. Variation of FAR with D/C Ratio.
Figure 3. Variation of FAR with D/C Ratio.
Infrastructures 09 00170 g003
Figure 4. MTTD trends in relation to D/C ratio for MFNN model in cross-validation.
Figure 4. MTTD trends in relation to D/C ratio for MFNN model in cross-validation.
Infrastructures 09 00170 g004
Table 1. Confusion Matrix for the Optimized MFNN Model using Cross-Validation.
Table 1. Confusion Matrix for the Optimized MFNN Model using Cross-Validation.
True NormalTrue IncidentClass Precision
pred. normal10,38155394.94%
pred. incident179340795.01%
class recall98.30%86.04%
Accuracy94.96%F-score90.30%
Table 2. Performance Comparison of the Developed AID Model with the Existing Literature.
Table 2. Performance Comparison of the Developed AID Model with the Existing Literature.
AID Model AuthorsDR (%)FAR (%)MTTD (min)
Developed model 95.961.010.89
SNDParkany and XIE [25] 921.31.1
SVM_LMotamed [38]870.074.3
SVM_RBMotamed [38]91.30.075.45
SVM_PMotamed [38]91.30.012.25
ANNMotamed [38]82.60.063.25
PNNMotamed [38] 95.60.33.84
Hybrid modelXIE et al. [39]97.30.061-
ANNCheu and Ritchie [87]801.54.95
GPS-based AIDD’Andrea and Marcelloni [30]91.68.37
IQD_SpeedAhuja [56]945.4-
IQD_Speed and OccupancyAhuja [56]924-
Decision TreeAhuja [56]973-
RFChakraborty et al. [61,132]973-
IQDZyryanov [5]974.812.4
ANNRossi et al. [106]97.6--
FLDogru and Subsa [99]93.090.4452.95
ANNDogru and Subsa [99]86.18-
RFDogru and Subsa [99]940.203-
SVMDogru and Subsa [99]884.2-
Video-based AIDRen et al. [121]96.60.721.16
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

ElSahly, O.; Abdelfatah, A. Developing a Machine-Learning-Based Automatic Incident Detection System for Traffic Safety: Promises and Limitations. Infrastructures 2024, 9, 170. https://doi.org/10.3390/infrastructures9100170

AMA Style

ElSahly O, Abdelfatah A. Developing a Machine-Learning-Based Automatic Incident Detection System for Traffic Safety: Promises and Limitations. Infrastructures. 2024; 9(10):170. https://doi.org/10.3390/infrastructures9100170

Chicago/Turabian Style

ElSahly, Osama, and Akmal Abdelfatah. 2024. "Developing a Machine-Learning-Based Automatic Incident Detection System for Traffic Safety: Promises and Limitations" Infrastructures 9, no. 10: 170. https://doi.org/10.3390/infrastructures9100170

APA Style

ElSahly, O., & Abdelfatah, A. (2024). Developing a Machine-Learning-Based Automatic Incident Detection System for Traffic Safety: Promises and Limitations. Infrastructures, 9(10), 170. https://doi.org/10.3390/infrastructures9100170

Article Metrics

Back to TopTop