To better understand the developed model performance, the confusion matrices of the cross-validation and testing processes are presented and discussed in detail. Additionally, a comparison is made between the performance of the developed model and some of the existing AID models. Finally, the analysis is carried out to examine the impacts of the factors considered in this research on the model’s efficacy.
4.2. Results of Cross-Validation Phase of Random Forest
In this section, the details of incident scenarios in the cross-validation dataset are presented. The calculations of DR, FAR, and MTTD of this dataset are illustrated. The confusion matrices presented in
Table 2 and
Table 3 from the cross-validation and testing phases require further analysis to calculate the three performance measures: DR, FAR, and MTTD. During both the cross-validation and testing phases, the model’s predictions are compared against the actual status of the traffic for each interval in both datasets. Based on this comparison, the model’s prediction is classified as true, false alarm, or missed. A sample of this comparison during one of the scenarios used in the cross-validation dataset is shown in
Table 4.
The observations made during this analysis are as follows:
Fluctuations in the incident alarm are observed after the incident had occurred (at the 30th interval) until the 70th interval (20 min) as illustrated in
Table 4. The algorithm sometimes labels intervals as normal despite them being incidents. This fluctuation arises from differences in volume, speed, and occupancy between upstream and downstream stations, which are input variables used for incident detection. Initially, significant differences between stations allow the model to detect incidents easily. However, as the incident persists, these differences diminish, potentially leading to incorrect labeling as a normal condition. Therefore, it is important to note that the detection is counted at the first alarm produced by the model, as once the operators receive an alarm, they will check the cameras to verify its authenticity.
In one of the scenarios, some false alarms occurred at consecutive intervals, as shown in
Table 5.
Treating these consecutive false alarms as separate alarms is unrealistic and affect the reliability of the model. An assumption regarding false alarms has been made in this paper to consider real-life conditions when calculating the FAR, consecutive false alarms are treated as a single false alarm if they persist for four intervals or less, representing real-life conditions. If there is a gap of more than four intervals, they are considered separate false alarms. This assumption aligns with operational scenarios and aids in estimating the real false alarms. When an incident alarm is received, its authenticity is verified, and appropriate action is taken. Keeping in mind that the interval is 30 s, the check of authenticity may consume more than one interval, therefore, consequtive alarms are considered as one alarm. It is important to note that this assumption may affect the detection time of an incident.
Based on these observations, it can be concluded that calculating DR and FAR from confusion matrices alone can be inaccurate. Further analysis is needed to obtain realistic DR, MTTD, and FAR values. Time to Detect (TTD) is calculated for each incident scenario, focusing on the first interval of incident detection. The DR and FAR are calculated by evaluating the successful incident detection and the number of false alarms, respectively. In this analysis, FAR is calculated as the ratio of the number of false alarms generated by the algorithm to the total number of times the algorithm was applied, which provides a more accurate measure. Analyzing these results provides a comprehensive understanding of the model’s performance.
The cross-validation dataset utilized in this analysis consists of 121 scenarios, comprising 22 normal conditions and 99 incident scenarios. It should be noted that the incident in each scenario starts at the 30th intervals and ends at the 70th interval (i.e., it lasts for 20 min) and the intervals before and after the incident are normal conditions. Accordingly, the total number intervals for incidents is 3960 while the number of intervals for normal conditions is 10,560. In
Table 6, the number of false alarms for both sets of scenarios is presented, alongside the number of detected incidents and the MTTD required to identify these incidents.
As shown in
Table 6, out of the 99 incident cases included in the dataset, the model is unable to detect 3 incidents (DR of 96.97%). The cross-validation dataset consists of 121 h of traffic data, with traffic state checked every 30 s, resulting in 14,520 model’s applications. During these applications, the model generated a total of 90 false alarms, which resulted in an FAR of 0.62%. In addition, the MTTD = 1.05 min. In order to provide a more comprehensive analysis of the model’s performance in detecting incidents of varying severity levels,
Table 7 summarizes its DR, FAR, and MTTD for each level of incident severity.
The results in
Table 7 reveal some interesting insights. The severity of an incident, measured by the number of lanes blocked, affects the model’s performance in detecting incidents. When the incident severity is medium to high (two or more lanes are blocked), the model consistently detects all incidents with a DR of 100%. However, incidents with minor severity (only one lane blocked) have a lower DR, possibly because their minor impact on traffic flow makes them less noticeable to the model, especially at low traffic volumes. The highest FAR value was observed in the scenario where four lanes were blocked, reaching 0.972. This indicates that the model is more prone to generating false alarms when a greater number of lanes are blocked. Conversely, the lowest FAR value of 0.613 was recorded when three lanes were blocked. In this scenario, the model demonstrated relatively better performance in detecting incidents, with fewer false alarms. However, there does not appear to be a clear increasing or decreasing trend in the FAR as the number of lanes blocked varies. The FAR values fluctuate across different incident scenarios, suggesting that the severity of incidents, as indicated by the number of blocked lanes, does not consistently correlate with the model’s performance in terms of false alarms. Regarding the MTTD, the MTTD decreases as the severity increases. Incidents with one lane blocked have the longest MTTD at 2.5 min. In contrast, incidents with four or five lanes blocked are detected much faster, with MTTD of 0.5 min and 0.565 min, respectively. This trend can be attributed to the fact that as the severity of the incident increases, its impact on various traffic parameters becomes more significant. Incidents involving the blocking of multiple lanes tend to create larger disruptions in the traffic flow, leading to more noticeable disturbances that can be detected and identified more quickly by the model. Hence, incidents with a higher severity level are detected at a faster rate, resulting in shorter MTTD values. It is important to consider that factors such as D/C ratios, spacing between detectors, and incident locations were not standardized or kept consistent across these incident severity scenarios. These variations in factors can contribute to the observed fluctuations in the FAR values. To obtain a more accurate understanding of the impact of incident severity on the model’s performance, a sensitivity analysis should be conducted. By controlling and standardizing the D/C ratios, detector spacing, and incident locations, the effects of the number of lanes blocked can be isolated, providing clearer insights into its direct influence on the model’s accuracy in detecting incidents without the confounding factors affecting the results.
Next, a sensitivity analysis is conducted to assess the impact of four factors, namely D/C ratio, incident severity, location of the incident, and distance between detectors, on DR, FAR, and MTTD. Each factor is analyzed individually to evaluate its effect on the performance metrics. It should be noted that the 0.4 and 0.5 D/C ratios use different blockage levels (2 or 4 lanes) and have constant incident location relative to the upstream detector. Also, the 0.9 and 1.1 D/C ratios are applied for different spacing between the detectors, different severity of the incident (2, 3, or 4 lane blockage), and different location of the incident. Therefore, these D/C ratios (0.4, 0.5, 0.9, and 1.1) are not considered in the sensitivity analysis. To ensure a fair comparison, incident scenarios with the same severity levels, locations, and detector spacings are selected for the analysis. Specifically, incident scenarios with D/C ratios of 0.6, 0.8, 1.0, and 1.2 are chosen for the analysis.
4.2.1. Impact of D/C Ratio on Performance Metrics
To investigate the impact of D/C ratio on the performance metrics of the incident detection model. The variation in DR, FAR, and MTTD, when the D/C ratio is varied between four ratios (0.6, 0.8, 1.0 and 1.2) is analyzed.
Figure 1 shows that the DR reached 100% for all D/C ratios, except for 0.6, where it reached 85.71%. The low DR at D/C of 0.6 can be attributed to the fact that three incidents out of the total are not detected by the model. These three incidents occurred at low traffic volume (D/C = 0.6), and one lane blockage out of the seven lanes of the road. Therefore, such minor incidents do not have significant effect on the traffic performance and are harder to detect, which resulted in a reduced DR of the model. To confirm this justification for low DR at D/C ratio of 0.6, the DR is calculated for all D/C ratios without the one lane blockage incidents that occurred at 0.6 D/C ratio, and it was consistently 100%. This suggests that the model’s performance can be influenced by the occurrence of minor incidents during low traffic volume, and excluding such incidents can lead to a significantly higher DR. Therefore, it is important to take into account the severity of incidents and the volume of traffic at the time of their occurrence, when assessing the performance of the model in detecting incidents.
Figure 2 shows that there is a variation in FAR with respect to D/C ratio. While the FAR values show some fluctuations, there is an overall trend of decreasing FAR with an increase in D/C ratio from 0.8 to 1.2.
To further understand the impact of minor incidents during low traffic volume on the FAR,
Figure 3 shows the relationship between FAR and D/C ratio excluding the one lane blockage incidents that occurred at 0.6 D/C ratio.
After excluding the minor incidents during low traffic volume, it becomes evident that the FAR decreases with the increase in D/C ratio. Specifically, the FAR remains relatively stable for D/C ratios of 1.0 and 1.2, with values of 0.606% and 0.635%, respectively. The decreasing trend of FAR with an increase in D/C ratio can be attributed to the fact that as the congestion level increases, vehicles are more likely to be closely spaced and moving slowly, which can result in more consistent behavior and patterns that are easier for the incident detection system to recognize. In contrast, in case of lower traffic volumes, vehicles may be more spaced and traveling at varying speeds, making it more difficult for the system to distinguish between normal and abnormal traffic behavior. This is consistent with earlier findings that suggested that the low traffic volume conditions lead to low occupancy values and a higher chance of false alarms [
121,
122].
Figure 4 shows that the MTTD values appear to increase as the D/C ratio increases from 0.8 to 1.2. This suggests that the model is more efficient at detecting incidents when the D/C ratio is low. One possible explanation for the increase in MTTD is that at higher traffic volume and congestion, there are queues at the blocked sections of the road, which can cause delays in traffic flow. As a result, incidents that occur may take longer time to create a detectable impact on the traffic performance, as the travel time between the detectors increases with congestion. Therefore, the detection of incidents can be further delayed.
Figure 5 shows the relationship between MTTD and D/C ratio without the one lane blockage incidents that occurred at 0.6 D/C ratio.
Figure 5 confirms that the inconsistent trend in
Figure 4 is mainly because of the existence of minor incidents during low traffic volume as shown in
Figure 5, the MTTD consistently increases with the increase in D/C ratio. It should be noted that
Figure 1,
Figure 2, and
Figure 4 included incidents with one lane blockage occurring at a D/C ratio of 0.6, which the model failed to detect some of them due to their minor impact during low traffic volumes. As a result, the trends of DR, FAR, and MTTD in these figures were not consistent. When excluding the minor incidents during low traffic volumes, the DR becomes 100% for all D/C ratios and
Figure 3 and
Figure 5 show a clearer trend for the FAR and MTTD. Therefore, these incident scenarios will be excluded in the upcoming analysis of the other factors to obtain a clear understanding of the impact of the factor of interest on the performance measures. These findings suggest that D/C ratio is an important factor to consider when evaluating the performance of traffic incident detection models.
4.2.2. Impact of Incident Severity on Performance Metrics
An analysis was conducted to examine the variation in DR, FAR, and MTTD as the lane blockage varied from 1 lane to 5 lanes blockage. The DR of the incident detection system shows a constant value of 100%, for all incident cases except for the one lane blockage at D/C ratio of 0.6, indicating that the system is highly accurate in detecting incidents with a significant impact on traffic flow. However, when considering incidents with one lane blockage at a D/C ratio of 0.6, the DR drops to 88.46%. This result is in line with the findings of previous research [
44,
81], indicating that the DR tends to be lower for minor severity incidents that occur during periods of low traffic volume.
The FAR values were recorded for different lane blockage categories: 1, 3, and 5 lanes. Interestingly, the FAR remained consistently low, staying below 1% across all lane blockage categories. The FAR was found to be 0.789%, 0.654%, and 0.699% for one, three and five lane blockages, respectively. The results indicate that the number of lanes blocked has a minor impact on FAR as there is a very small variation. These results suggest that the incident detection system is not significantly impacted by the incident severity, except for minor incidents during low traffic volume. The lower FAR values across all lane blockage categories are a good indicator of the efficacy of the detection model in distinguishing actual incidents from normal traffic conditions. It should be noted that when the incidents with one lane blockage at D/C of 0.6 are included, the FAR is 0.64, which is lower than the case without such incidents. The system did not create any false alarms nor detect any incidents during such scenarios. Therefore, when these scenarios were excluded, the FAR increased as the number of false alarms is compared to a smaller number of application intervals.
However, when considering the MTTD in relation to the number of lanes blocked, a different trend was observed. The results revealed a notable variation in MTTD with respect to the number of lanes blocked. When one lane was blocked, the MTTD was recorded as 2.37 min. However, as the number of blocked lanes increased to three and five, the MTTD significantly decreased to 0.589 min and 0.565 min, respectively. These findings indicate that as the incident severity increases, the MTTD decreases, suggesting a faster detection time for incidents. This could be due to the fact that incidents with higher severity are easier to detect as they have a higher impact on the traffic flow, allowing the system to detect such incidents in a shorter time. The current findings are consistent with the results reported by Cheu and Ritchie [
44].
These results suggest that the severity of an incident, as represented by the number of blocked lanes, has a significant impact on the MTTD. However, it has no impact on the detectability of the incidents and a minor impact on FAR. The DR of the system is consistently high for all lane blockage, except for one lane blockage at D/C of 0.6. The FAR was marginally affected as the number of lanes blocked changes. The MTTD of the system decreases as the severity of the incident increases, indicating that the system is faster at detecting severe incidents with higher lane blockage.
4.2.3. Impact of Incident Location on Performance Metrics
The incident location varies to be 0.25, 0.5, and 0.75 of the distance between the detectors, from the upstream detector. The DR remained 100% for all three incident locations, indicating the system’s high accuracy in detecting severe incidents that have a significant impact on traffic flow. It should be noted that this analysis excludes one lane blockage incidents at 0.6 D/C ratio, due to their minor influence during low traffic volume, as mentioned previously. It is worth noting that when the incidents with one lane blockage at D/C of 0.6 are included, the DR values are 93.1%, 94.12% and 100% for 0.25, 0.5 and 0.75 incident locations respectively. Regarding the FAR, the FAR decreases as the incident moves further from the upstream detector station, with values of 0.99, 0.74, and 0.36 observed for incident position ratios of 0.25, 0.5, and 0.75, respectively. With the inclusion of incidents that involve a single lane blockage at D/C of 0.6, the FAR values for incident locations of 0.25, 0.5, and 0.75 are 0.99%, 0.71%, and 0.36%. It can be noted that when one lane blockage incidents scenarios were excluded, the FAR slightly increased as the number of false alarms was compared to a smaller number of application intervals. The decrease in FAR can be attributed to the fact that the model did not generate any alarms (false or true) during the minor incidents, resulting in a reduction in the FAR. On the other hand, the MTTD values showed a clear correlation with the incident’s location. As the incident is moved further from the upstream station, specifically from 0.25 to 0.75 of the distance between the detectors, an increase in the MTTD is observed. The MTTD values increase from 0.937 min to 1.522 min. This can be attributed to the fact that as the incident moves further away, the time it takes for its effects to propagate to the upstream detector increases. As a result, the detection of the incident is delayed, leading to higher MTTD values.
It can be concluded that the system is highly effective in detecting incidents at all locations, with a DR of 100%. The FAR is slightly impacted by the incident location, while the MTTD of the system is consistently increasing, as the incident location moves further away from the upstream detector. This trend indicates that the incident location relative to the detector plays a vital role in the MTTD, with incidents occurring closer to the upstream detector being detected faster. Neglecting incident location as a factor can compromise the efficacy of detection algorithms and undermine their overall effectiveness.
4.2.4. Impact of Distance between Detectors on Performance Metrics
The distance between the two detectors varies from 500 to 1500 m, and the DR, FAR, and MTTD are evaluated to assess the impact of this variation on the model’s performance. The results show that DR is constant at 100% for all three spacing values, 500 m, 1000 m, and 1500 m excluding minor one lane blockage incidents at 0.6 D/C ratio. The inclusion of such incidents leads to 100% DR value for 500 m spacing and a reduction in DR values to 96.97% and 92.86% for detector distances of 1000 m and 1500 m, respectively. Notably, of the three minor incidents at 0.6 D/C ratio that were not detected, one occurred at 1000 m detector spacing and the remaining two occurred at 1500 m spacing. An analysis of the relationship between the FAR and the distance between detectors revealed that the FAR was recorded as 0.432%, when the distance between detectors was 500 m. As the distance between detectors increased to 1000 m, the FAR showed an increase to 0.673%. Furthermore, when the distance was further extended to 1500 m, the FAR value reached 1.03%. These findings demonstrate a clear trend in the FAR values, indicating that as the distance between detectors increases, the FAR also increases. The increase of the FAR with the increase of the distance between detectors may be due to the longer travel time of vehicles between detectors. Longer travel time can cause fluctuations in traffic measurements, such as volume, speed, and occupancy, at both the upstream and downstream detectors. These fluctuations may lead to variations in the speed measurements. These fluctuations in traffic measurements can cause false triggers of incident alarms. Regarding the MTTD, it was noted that larger detector spacings are associated with longer detection times. The MTTD was recorded as 0.593 min, 0.788 min, and 1.7 min for the three spacings, respectively. This can be attributed to the increased travel time of vehicles between the detectors, as mentioned previously, which can cause delays in detecting incidents. This is aligned with the findings of Rossi et al. [
81] who reported that an increase in the distance between detectors results in higher MTTD values.
The overall observations from this analysis indicate that smaller detector spacings result in better performance in terms of FAR and MTTD. This is due to the closer proximity of the detectors, which reduces the detection time and traffic measurement fluctuations between the detectors. The DR remains 100% in all cases, indicating that the system is capable of detecting all incidents regardless of the distance between the detectors. However, while smaller detector spacings are more effective in detecting incidents, they may not be practical due to installation and maintenance costs. On the other hand, larger detector spacings may be more cost-effective and easier to maintain but may result in higher FAR and MTTD values, which is consistent with previously reported results [
84]. Therefore, the decision to select an appropriate detector spacing should consider specific application requirements, available resources, and trade-offs between the performance (in terms of DR, FAR, and MTTD) and the associated costs. An optimal balance between these factors is necessary to ensure an effective and cost-efficient incident detection system.
The DR, TTD, and number of false alarms of each scenario in the cross-validation dataset is calculated, and the results are presented in another publication [
123].
4.4. Discussion of Random Forest Model Results
After analyzing the impact of each of the four factors individually, the impacts of all four factors together on the detectability, TTD, and FAR of incidents in the cross-validation and testing phases is discussed. A selection of incident scenarios that are not detected or have relatively long TTD are analyzed to understand the combined impact of all four factors in the cross-validation and testing phases. During the cross-validation phase, it was previously noted that the model failed to detect three incidents that blocked one lane and occurred during low traffic volume at a D/C ratio of 0.6. These incidents took place at a detector spacing of 1000 and 1500 m respectively. The model’s inability to detect these incidents can be attributed to three key factors: traffic volume, incident severity, and detector spacing, which have been discussed earlier. The low severity of these incidents, combined with their occurrence during low traffic volume and the significant distance between the detectors, all contributed to the model’s inability to detect them. The model was able to detect an incident where only one lane was blocked, and the D/C is 0.6. The incident occurred 750 m from the upstream station with a spacing of 1000 m between the stations. However, the model detected the incident 9.5 min after its occurrence, which is not a good indicator of the detection performance. This delay in detection can be attributed to the low severity of the incident, the low D/C ratio, and the distance between the detectors. These factors affected the model’s ability to detect the incident in a timely manner. This finding supports the earlier conclusion that minor severity incidents during low traffic volume can be hard to detect or undetectable. This observation is consistent with the findings of Cheu and Ritchie [
71], Ahmed and Hawas [
121] and Rossi et al. [
81] who reported that incidents involving minor blockage during low traffic volume are associated with low DR and high MTTD values. At a D/C ratio of 1.2, two incidents with high TTD were observed during the cross-validation phase. These two incidents are minor one lane blockage incidents located at distances of 125 and 1125 m from the upstream station, respectively. The distances between the detector stations for these two incidents are 500 and 1500 m, respectively. The model took 6 and 18 min, respectively, to detect them. The extended TTD of the two incidents during the high D/C ratio of 1.2 can be attributed to a combination of factors. Firstly, the severity of the incidents played a role in the delay of detection. The congested conditions during a D/C ratio of 1.2 may have also contributed to the long MTTD, as illustrated in
Figure 5. During periods of high demand, such as a D/C ratio of 1.2, the road network experiences heavy traffic congestion. This results in more vehicles on the road, which tend to travel in platoons or sometimes wait in queues, due to the lack of available space. The existence of these platoons can cause slower travel speeds, which in turn delays the incident detection by the model. Moreover, the presence of platoons makes it more difficult for the model to detect changes in traffic patterns that could indicate an incident.
Likewise, during the testing phase, a severe five lane blockage incident occurred at 1.2 D/C ratio, and it occurred at 125 m from the upstream station, with a spacing of 500 m between stations. The TTD for this incident is 15 min, which is high. Such high value of TTD can be attributed to the oversaturated condition of the road network, as discussed earlier. In this incident, the demand was already above the road’s capacity by 20%, and the vehicles on the roads were already queued due to congestion. Therefore, when this severe incident occurred, it further exacerbated the congestion, resulting in a further buildup of queued vehicles. However, the impact of this incident would take some time until the queues of vehicles ahead of the incident location are discharged, so that the impact of this incident reached the downstream station. Therefore, the combination of a high D/C ratio, the severity of the incident and the incident location resulted in a long TTD of 15 min. Notably, during the evaluation of the incident detection system, it was observed that false alarms were produced in certain instances. Most of these false alarms were generated in the 1–2 min range post-incident termination. This is due to the residual impact of the cleared incident, which can cause fluctuations in traffic performance that may be misidentified as a persistent issue by the model. In such cases, the operator can adopt a wait-and-see approach, allowing for a brief time to assess whether the situation is truly an ongoing incident or just a residual effect. This approach can help mitigate the occurrence of false alarms, ensuring the efficiency and accuracy of the incident detection system. In conclusion, it is highlighted by the analysis of the cross-validation and testing phases that considering various factors such as severity of the incident, traffic conditions, and location and spacing between the stations is crucial. The detectability and MTTD of incidents can be significantly influenced by these factors individually or together, ultimately impacting the accuracy of the system’s ability to detect them. Thus, the potential of the incident detection system to be a useful tool for real-world incident detection is demonstrated by the results.
4.5. Comparison of the RF Model with Other AID Systems
This section aims to compare the performance of the developed model with some of the AID models identified in the literature review that exhibit high performance. The effectiveness of the model is evaluated and compared using key performance metrics such as DR, FAR, and MTTD. The DR, FAR, and MTTD reported in the literature for the selected AID model are summarized in
Figure 6,
Figure 7 and
Figure 8.
It can be noted from
Figure 6,
Figure 7 and
Figure 8 that the developed RF model achieved high performance in terms of DR, FAR and MTTD and it has superior performance than the majority of the models in the literature. However, some AID models achieved better performance in terms of one or more of the performance metrics in comparison with the developed model. However, these models have some limitations and weaknesses. The hybrid AID method developed by Xie et al. [
29] achieved high DR and low FAR on the I-205 and I-880 datasets. Specifically, the framework achieved a DR of 97.3% and FAR of 0.061% on the I-205 dataset, and a DR of 94.7% and FAR of 0% on the I-880 dataset. The average DR for the two cases is lower than the model developed in this research. In addition, a weakness of this framework is that the MTTD was not measured, which could limit the understanding of the model’s effectiveness in real-world scenarios. Another potential limitation is that the framework used an oversampling technique to generate synthetic incident samples. The oversampling technique can potentially limit the real-world applicability of the developed model. The reason for this is that the synthetic samples generated do not represent actual traffic incidents, which can lead to false positives or overestimation of the incident detection performance when the model is applied to real-world scenarios. In other words, the model’s performance in real-time application may differ from its performance on synthetic data generated through oversampling. Another potential limitation of the oversampling approach is overfitting, which can occur when a model is trained on a limited set of data. This may lead to a good model performiance on the training data but a poorl performance on new and unseen data. Furthermore, the use of synthetic incident samples to balance the database between incident and non-incident cases may not be realistic. Therefore, balancing the incident and non-incident cases in the database may not be necessary, and may even introduce biases in the model’s training data. The decision tree and RF developed by Ahuja [
40] achieved high DR and reasonable FAR but they didn’t measure the MTTD of these models. While DR and FAR are important metrics to evaluate the performance of an AID system, they only provide a partial picture of how well the system performs in real-world scenarios. In addition to detecting incidents accurately and minimizing false alarms, it is also crucial for an AID system to detect incidents quickly, as the timely detection of incidents can help reduce their impact on traffic flow and prevent secondary accidents. Likewise, The ANN model developed by Zyryanov [
75] accomplished a slightly higher DR than the developed model in this paper during training phase. Nonetheless, the evaluation of the model’s performance only considered DR and did not include FAR and MTTD, which are critical metrics in assessing the model’s reliability and efficiency. As a result, the evaluation may not provide a comprehensive understanding of the model’s effectiveness, limiting its practical applications. On the other hand, the FL system developed by Rossi et al. [
81] has lower FAR than the developed RF model but it has lower DR and higher MTTD. Additionally, the training and testing of the system were based on simulated traffic data, and the study only tested the system’s performance for a limited range of flow rate values and distances among detector stations without considering other factors. The RF model created by Dogru and Subasi [
70] achieved lower FAR of 0.2% but it has lower DR of 94% and the MTTD of this model was not reported. Besides, their model have limited practical applicability due to several assumptions. For example, the assumption that only probe vehicles are equipped with V2V communication devices and that they broadcast their position and speed every second. Additionally, the models assume that probe vehicles can calculate the position of transferring vehicles using signal processing and antenna techniques and that they can aggregate microscopic traffic values for each vehicle over the last 10 s. These assumptions may not hold in real-world scenarios and thus limit the models’ practical use. The video-based detecting and positioning method developed by Ren el al. [
95] exhibited DR, FAR, and MTTD values that were in close proximity to those achieved by the developed model. In spite of that, the lighting conditions, extreme weather conditions and coverage range of the camera that is used to capture traffic video are the main limitations that can impact the performance of this method. In addition, the method requires a significant number of computational resources and may not be suitable for real-time applications or for use in areas with limited computing power. Moreover, they used oversampling technique to generate synthetic incident sample to balance between incident and non-incident samples in the datasets used to train and test the model. As mentioned earlier, using oversampling technique to generate artificial incidents is not realistic.
In summary, the comparison of the developed RF model in this paper with other AID systems showed that the model has better performance than most of the existing ones. The developed RF model considered a wider range of factors that have significant impacts on AID performance. Compared to previous models, the developed model integrated more factors, making it more comprehensive and generic. The aim of the developed model is not only to detect incidents accurately and rapidly, but also to establish effective and generic incident detection model for freeways. Moreover, the use of ML techniques such as RF enabled the model to learn from data and adapt to changing traffic conditions, making them more robust and adaptable to new scenarios. This ensures the generalizability of the model to new data and the potential for future applications in real-time traffic management systems. Overall, the developed RF model offers a promising solution for accurate and effective incident detection on freeways.