Road Accident Analysis and Prevention Using Autonomous Vehicles with Application for Montreal

Abstract

Road safety in cities is becoming a bigger concern worldwide. As more people own cars and traffic congestion increases on old roads, the risk of accidents also grows, which severely affects victims and their families. In 2023, data from the Société de l’Assurance Automobile du Québec (SAAQ) reported that 380 people died in traffic accidents in Quebec. A study of road accidents in Montreal between 2012 and 2021 looked at the most dangerous locations, times, and traffic patterns. In this paper, we investigate the role of autonomous vehicles (AVs) vs human-driven vehicles (HDVs) in reducing road accidents in mixed traffic situations. The reaction time of human drivers to road accidents at signalized intersections affects safety and is used to compare the difference between the two situations. Microscopic traffic simulation models (MTMs) namely the Krauss car-following model is developed using SUMO to assess the vehicles performance. Case study 1 assesses the effect of reaction time on human-driven vehicles. The findings show that longer reaction times lead to more collisions. Case study 2 looks at autonomous vehicles and how human-driven vehicles interact in mixed traffic. The simulations tested various levels of AV penetration (0%, 25%, 50%, 75%, and 100%) in mixed traffic and found that more AVs on the road improve safety and reduce the number of accidents.

1. Introduction

Traffic accidents are one of the leading causes of death worldwide and put a heavy strain on the global economy. With millions of people killed or injured each year, the rising number of accidents is a major public health issue. The 2023 Global Status Report on Road Safety by the World Health Organization (WHO) [1,2,3] notes that efforts to improve road safety have had only a small positive effect. While road traffic deaths have slightly decreased to 1.19 million annually, this is still a significant number, affecting many families. The report calls for coordinated global efforts to reduce traffic fatalities by 50% by 2030.

A collision occurs when a vehicle hits another car, a person, an animal, or an obstacle. Traffic accidents negatively affect society, as they cause emotional and psychological harm to the victims and their families. These families often experience long-term mental health issues. A recent study by Salman Zafar [4] for OECD countries looked at road safety by measuring the number of fatalities per 100,000 people. It found that the USA had the highest number of fatalities, followed by New Zealand and Canada, while the UK had the lowest. Although OECD countries are generally wealthier and more industrialized, they face significant road safety concerns due to having more vehicles than non-OECD countries. The study shows that Canada’s fatality rate has improved over time, dropping from 7.18 to 4.59, reflecting better road safety measures, though the loss is still considerable. The study also found that men have a higher road death rate than women, mainly because men tend to drive more often and longer distances, engage in riskier driving behaviors (like speeding and driving under the influence), and are less likely to wear seat belts [5].

Several factors have contributed to the increase in road fatalities, especially in Canada, including the pandemic lockdown, which led to a rise in risky driving behaviors like impaired driving and speeding. There has also been a noticeable increase in motor vehicle and bicycle fatalities compared to previous years, with accidents in rural areas proving to be more deadly than those in urban areas [6]. In Quebec, there were 4.5 fatalities per 100,000 people and 4.7 fatalities per billion vehicle kilometers traveled (VKT) [6]. A comparison of road fatalities highlights differences and similarities across regions, helping to identify areas that need targeted road safety efforts. The UK had the lowest number of fatalities, likely due to their strict road safety measures, such as regular safety audits, identifying high-risk roads, maintaining road infrastructure, enforcing tough traffic laws, imposing heavy penalties, raising public awareness, and strictly punishing unsafe driving practices [7].

The Road Safety Record for Quebec [8] shows the trends in road accidents. In 2023, 380 people died in traffic accidents, which represents a 4.5% decrease from 2022, but a 7.5% increase compared to the average from 2018 to 2022. According to Figure 1, most of the deaths involved occupants of automobiles and light trucks, with 234 fatalities—an increase from previous years. Motorcyclists accounted for 54 deaths, which was lower than both the 2022 count and the 2018–2022 average (Société de l’Assurance Automobile du Québec, 2024). While fatalities among heavy truck occupants decreased, the number of deaths among light vehicle occupants was higher compared to previous years.

The number of road fatalities also varied across different age groups in 2023. Deaths decreased among individuals aged 0 to 14, 25–34, 55–64, and 65–74, while there was an increase in fatalities for people aged 15–24, 35–44, and 45–54. The 15–24 age group saw the most significant rise, with 59 deaths, marking a 20.4% increase compared to 2022 and a 31.7% increase compared to the 2018–2022 average. This demographic accounted for 15.7% of total road deaths. On the other hand, fatalities among individuals aged 75 and older dropped by 6.7% from 2022, though they remained 10.7% higher than the five-year average, as shown in Figure 2. This shows that road traffic accidents are a serious problem in Quebec, Canada.

Figure 3 illustrates that human factors are the primary cause of fatal accidents, with speeding, impairment, distraction, and other hazardous driving behaviors leading the way. While environmental and vehicle-related factors do contribute, this data suggests that behavior-change strategies, especially through the adoption of advanced driver assistance systems (ADASs), could play a crucial role in reducing future fatal collisions.

Given the limitations of human drivers and the high frequency of road accidents in Montreal, there is an urgent need for innovative solutions to enhance road safety and minimize accident-related injuries and fatalities. One of the most promising advancements in this area is the implementation of autonomous vehicles (AVs) [9,10]. AVs are equipped with cutting-edge sensing technologies, real-time data processing, and machine learning capabilities, which enable them to perceive and respond to road conditions more effectively than human drivers [10]. Multiple studies have shown that AVs have the potential to significantly reduce accidents caused by human error or negligence [11,12,13].

In this paper, we address the problem of road traffic accidents in Montreal and analyze the role of autonomous vehicles (AVs) in mitigating them. Autonomous vehicles are a rapidly developing technology, and their widespread adoption is still in progress. While there has been considerable research on AVs and their impact on road safety, studies specifically examining the effect of driver reaction time in relation to road collisions, particularly in Montreal, are scarce. This gap presents an opportunity for further exploration in this evolving and high-stakes field. The key contributions of this paper include:

• Effect of driver reaction time in road accidents: The role of human driver reaction time, and how it influences accident frequency and severity in comparison to AVs’ faster reaction capabilities, has not been thoroughly investigated.
• Limited localized studies on Montreal’s urban traffic risks: Most existing research focuses on broader traffic patterns, without addressing the specific risks and factors unique to Montreal’s urban road environment.
• Lack of in-depth analysis of road accidents using microscopic simulation and car-following models: While some studies use macroscopic simulations, a detailed microscopic approach, such as the use of car-following models to simulate individual vehicle behavior and interactions, is underexplored.
• Limited study on HDV–AV interaction in mixed traffic scenarios in Montreal: The interaction between human-driven vehicles (HDVs) and autonomous vehicles (AVs) in Montreal’s mixed traffic environment remains poorly understood, especially with respect to accident risks and safety dynamics.

The rest of the sections are organized as follows. Section 2 presents the problem statement. The literature review is covered in Section 3. Section 4 and Section 5 present the methodology and case study, respectively. Lastly, we conclude with directions for future works in Section 6.

2. Problem Statement

Road safety remains a critical public health concern globally, with developed nations like Canada facing significant challenges. Montreal, in particular, reports a higher incidence of road accidents compared to other major Canadian cities. In 2019 alone, the city witnessed over 20,000 road accidents, many of which resulted in injuries or fatalities [14]. Despite significant advancements in vehicle safety technologies and traffic management systems, accidents on the streets continue to pose a substantial threat to public health and safety [15]. The growing presence of human-driven vehicles (HDVs) exacerbates this problem, leading to increased traffic congestion, human errors, and heightened safety risks [16,17]. This study aims to analyze Montreal’s road accident dataset and explore the potential role of autonomous vehicles (AVs) in mitigating traffic-related issues and improving road safety in the city.

This research focuses on analyzing road accident data within Montreal to gain insights into factors like high-risk locations and accident-prone times of the day. The objective is to examine how human driving behavior contributes to accidents, particularly the risks posed by distracted drivers. Additionally, the study explores whether faster-reacting vehicles, such as autonomous vehicles (AVs), could be a potential solution to improve road safety.

3. Literature Review

Road accidents are a major contributor to injuries and fatalities globally. To mitigate the frequency and severity of crashes, effective analysis and prevention strategies are crucial. In this section, we provide an in-depth overview of the methods used to analyze road accidents and the strategies employed to prevent them.

3.1. Road Accident Analysis

Existing approaches for road accident analysis can be categorized into:

• Statistical analysis [18,19,20]
• Geospatial analysis [21]
• Machine learning and AI techniques [22,23,24,25]
• Crash reconstruction [26,27,28,29]
• Human factors and behavioral analysis [30,31,32,33]
• Traffic flow and congestion analysis [34,35,36,37,38]

Statistical analysis involves application of techniques such as descriptive statistics (e.g., accident counts and severity), time series analysis (e.g., monthly trends), and regression models (e.g., identifying risk factors like speed and time of day). to identify trends, frequency, and contributing factors in historical crash data. These methods are typically used to assess high-risk hours or demographics. Studies have also explored the impact of sleepiness [39,40], gender [41], weather conditions [42], and glare [43] on road safety. Examples of historical crash data sources include NHTSA (USA), IRAD (India), and local data.

Geospatial analysis (GIS) is employed to visualize and analyze crash locations. Techniques like heatmaps, spatial clustering (e.g., DBSCAN, K-means), and kernel density estimation (KDE) are commonly used. GIS software such as ArcGIS and QGIS have been used to help in identifying accident-prone areas (black spots [44,45,46]). Several reviews have highlighted the role of spatial approaches in road safety.

Machine learning and artificial intelligence techniques are increasingly used to predict and classify accident likelihood and severity. Examples of machine learning techniques are classification algorithms (e.g., random forest, neural networks), clustering (e.g., accident type segmentation), and predictive analytics (e.g., crash probability). Frameworks like Scikit-learn, TensorFlow, and PyTorch v2.7.1 are commonly used for predictive modeling. Machine learning methods also facilitate real-time accident prediction based on traffic patterns. Researchers have applied ML models for tasks like driving style classification [47,48,49], hazard scenario identification in non-motorized transport, and distraction studies related to motor vehicle accidents. Some studies have also explored weather-related road safety using logistic regression or traffic accident prediction using regression models [50,51]. Deep neural networks have been utilized to create vehicle dynamics models that predict real-time driving distance and velocity, which can be valuable for autonomous vehicles [52]. Machine learning approaches however face challenges with understanding and interpreting large datasets.

Crash reconstruction aims to recreate accident scenarios to analyze the causes and mechanics behind crashes using simulation software (e.g., PC-Crash, VISSIM) and physics-based calculations. This method is widely used in forensic analysis, legal proceedings, and insurance investigations. Some studies propose frameworks for proactive crash risk prediction, especially for behaviors like lane changes.

Human factors and behavioral analysis focuses on driver-related causes such as fatigue, distraction, and decision-making. Techniques like eye-tracking, simulators, reaction time studies, and observational surveys are used in this analysis. The insights gained from these studies contribute to policy design, educational programs, and traffic safety initiatives.

Traffic flow and congestion analysis examines traffic dynamics that contribute to accidents. This includes traffic simulations (both microscopic and macroscopic), queueing theory, and flow-density relationships. Simulation tools like VISSIM, AIMSUN, PC-Crash, and SUMO [53,54,55] are commonly used for modeling traffic and accidents. There are three primary types of traffic models: macroscopic (for network-wide traffic flow), microscopic (for individual vehicle behavior), and mesoscopic (a balance between accuracy and computational efficiency). Microscopic models are highly detailed but computationally expensive, making them less suitable for large-scale metropolitan simulations. However, they are ideal for applications like signal optimization, congestion management, and road accident prevention.

3.2. Road Accident Prevention

Car accidents worldwide are influenced by a variety of factors [56], including traffic conditions and weather patterns [57,58], substance use (such as drugs and cannabis) [59,60], driver fatigue [61], aggressive driving behaviors [62], speed [63], and even passenger seating positions [64]. Other contributing elements include inattention [65,66,67], angle collisions [68], and environmental conditions like precipitation [69]. The road accident prevention methods can be classified into:

• engineering and infrastructure solutions,
• enforcement and regulatory measures,
• education and public awareness,
• vehicle safety technology, and
• real-time monitoring and telematics.

Engineering and infrastructure solutions include road design improvements (roundabouts, elevated pedestrian crossings, medians), signage and lighting (reflective signs, improved street lighting), and smart infrastructure (intelligent traffic signals, real-time warning systems). Eboli and Forciniti [70] study a smartphone based system for sustainability and transportation safety. Several studies recommend automated vehicles for safety [71,72,73,74,75,76,77].

Enforcement and regulatory measures include traffic laws and penalties (strict enforcement of speed limits, helmet laws, seat belts), automated enforcement (red light cameras, speed cameras), alcohol and drug testing (random checks, zero-tolerance policies).

Education and public awareness include campaigns and workshops (awareness on drunk driving, speeding [78,79], use of mobile phones), driver training programs (defensive driving, simulation-based courses), targeted awareness (young drivers, motorcyclists, elderly drivers).

Vehicle safety technology includes advanced driver assistance systems (ADAS) (lane keeping assist, emergency braking, adaptive cruise control), event data recorders (Black Boxes) (for post-crash analysis and behavioral monitoring), Vehicle-to-Everything (V2X) (communication between vehicles, infrastructure, and pedestrians) [80].

Liang et al. [81] study presents an autonomous driving framework to systematically enhance AVs’ safe driving capabilities in high-speed cruising scenarios by integrating behavioral decision-making [82], path-planning, and motion-control modules. Their simulation results show that the proposed integrated framework may successfully direct AVs to achieve high-speed cruising while avoiding crashes. The proposed strategy improves the average reward in test scenarios by 10.25% over the typical sequential framework. Fagnant and Kockelman [83] study opportunities, barriers and policy recommendations for autonomous vehicles.

Real-time monitoring and telematics include connected vehicles (sharing live traffic and hazard data), telematics (monitoring driving behavior (speed, harsh braking)), smart traffic management centers (dynamic rerouting, real-time incident detection). Boyle and Mannering [84] study the impact of traveler advisory systems on driving speed. The simulation findings shows AVs to achieve high-speed cruising while avoiding crashes.

In this paper, we are exploring the role of autonomous vehicles in mitigating road traffic accidents in the context of City of Montreal. First, we analyze the road accident data for Montreal. Descriptive statistics and microscopic traffic simulation are then used for modeling reaction times to road traffic accidents analysis in the context of Montreal.

4. Research Methodology

The proposed methodology comprises of two steps. In step 1, we perform road accident data analysis and in the second step, we evaluate the potential of AVs in minimizing road accidents.

4.1. Road Accident Data Analysis

The Road Collision dataset (Collisions Routières) [85], which spans traffic accident statistics from 2012 to 2021 in Montreal was used. It has been instrumental in understanding the various factors contributing to accidents in the city, including their locations, environmental conditions, and time of occurrence. At the national level, 67% of fatal accidents in 2022 were linked to human factors such as impaired driving, speeding, and distractions [6]. This insight into human error as a major cause of accidents led us to focus on examining the relationship between reaction time and the frequency of accidents.

Following the accident analysis, we adopted microscopic traffic simulation methodology to study how human behavior influences traffic dynamics.

This approach captures individual vehicle behavior, enables realistic intersection and signal modeling, facilitates collision analysis, and assesses the performance of AVs in mixed traffic scenarios. SUMO (Simulation of Urban Mobility) software was utilized to simulate various scenarios, providing a clearer understanding of how driver reaction time impacts road safety.

4.1.1. SUMO Driving Behavior

SUMO version 1.8.0 [86] allows for the simulation of heterogenous traffic for AVs along with HDVs to interact. This feature is essential for assessing how AV penetration affects general traffic efficiency and safety.

In SUMO, the lane-changing and car-following models are independent modules. The Krauss car-following model is used by default for longitudinal dynamics in SUMO, while the SL2015 lane-changing model controls lateral behavior, particularly when sublane simulation (with a defined lateral resolution) is used. This combination allows vehicles to maintain safe, collision-free longitudinal behavior (Krauss) while executing realistic lane changes (via SL2015). To examine how these cars affect one another in different traffic situations, various model parameters (such as aggression, cooperative behavior, and reaction time) are adjusted.

4.1.2. Krauss Car-Following Model

The Krauss model, known for its collision-free dynamics and simplicity, underpins the longitudinal control of vehicles within the simulation. In this study, vehicle behavior is simulated using SUMO’s freeway driving mode, which leverages the Krauss car-following model in capturing realistic changes in speed, and safe gap maintenance on freeways. For Automated vehicles (AVs), an AV driving behavior is implemented using the same Krauss framework; however, these vehicles are assigned a distinct set of parameter values that reflect their enhanced driving capabilities. These parameters are based on the SUMO default that allow the AVs to exhibit more advanced maneuvering characteristics while still ensuring safe operation.

In SUMO, the interactions of vehicles with the driver or other vehicles is established through differential equations in which one’s behavior is modeled using longitudinal (car-following) or lateral (lane-changing) modeling behaviors, with the former being where car speed is conditioned on the vehicles ahead [87].

The following equation shows the Krauss car model:

• Safe Speed Calculation: Vehicles adjust speed to avoid collisions with the car in front.
• Acceleration and Braking: Vehicles accelerate until the maximum speed limit unless constrained by a leading vehicle.
• Time Step-based Updates: Vehicle’s speed is updated at every step during simulation.

$v_{desired}=min(v\left(t\right)+a\ast \Delta t,v_{safe},v_{max})$

• $v\left(t\right)$: current velocity
• $\mathrm{\Delta }t$: time step
• a = acceleration
• $v_{max}$: maximum speed limit of the vehicle
• $v_{safe}$: maximum safe velocity considering the leading vehicle.

4.1.3. Lane Changing Behavior

The SL2015 lane-changing model in SUMO is intended for high lateral resolution (or sublane simulation) and is used in combination with the car-following Krauss model to simulate realistic vehicle behavior. Lane change parameters control a vehicle’s lateral movements when they proceed to the adjacent lanes. Similar to other parameters, the autonomous vehicles have more rigorously selected values compared to conventional vehicles.

In this model, vehicles constantly assess the neighboring traffic taking into account the gaps present and the relative speeds of adjacent vehicles to determine whether a lane change is useful. Some of its most important behaviors are:

• Dynamic Gap Acceptance: Cars utilize parameters such as lcStrategic and lcSpeedGainLookahead to evaluate the space available in neighboring lanes and to forecast impending slowdowns or obstructions. This assists in deciding whether a lane change will result in a considerable speed gain.
• Cooperative Behavior: The model includes cooperative features (through parameters like lcCooperative and lcCooperativeSpeed) whereby vehicles not only determine when to switch lanes but also coordinate their speeds. This cooperative behavior ensures that the lane change can be carried out without disturbing the nearby traffic excessively.
• Lateral Dynamics Control: SL2015 regulates the precise lateral movement according to lateral acceleration limits and sublane positioning parameters (e.g., lcSublane and lcMaxSpeedLatStanding). This prevents “sliding” or aggressive lateral movement, making the vehicle smoothly switch lanes.

Overall, the SL2015 model enhances simulation realism through the ability of vehicles to make nuanced lane-changing decisions both in terms of safety and efficiency. Its integration with the Krauss car-following model ensures that when vehicles execute lateral maneuvers, they maintain safe longitudinal dynamics, resulting in a more realistic simulation of autonomous and human-driven traffic flows.

5. Application Results

In this section, we present the results under two categories. Firstly, the accident data analysis results for the city of Montreal are covered. Next, we conduct microscopic traffic simulation and scenario analysis to model the severity of the problem and present the results for AVs as a potential solution for mitigating road traffic accidents via simulation.

5.1. Data Overview

“Données Québec” is an open data portal created in collaboration between cities and the Government of Quebec. The volume of traffic on the roads of Montreal was collected from the “donneesquebec.ca” website. Real-time traffic flow was analyzed on roads and intersections at peak hours for simulation scenarios using dataset “Comptages Des Véhicules, Cyclistes Et Piétons Aux Intersections Munies De Feux De Circulation” [88]. The Collisions Routières dataset [85] consists of collision data for Montreal between 2012 and 2021.

The Road Collision dataset contains information on all accidents that occurred in Montreal, including those with material damage only (MDS) of less than USD 2000 and recorded by Service de police de la ville de Montréal (SPVM) for analysis based on geolocation. This dataset excludes collisions that took place on the roadway network. It has 68 attributes and 218,272 rows, which provide insights regarding the severity, location of accidents, weather conditions, vehicle types, timeline of accidents, and various other details. Fifty-three of these attributes are numerical variables (29 int, 24 float) and fifteen attributes are categorical variables (type object). We focused on the following attributes of collisions:

• DT_ACCDN: Date of collision.
• HR_ACCDN: Hour of collision.
• JR_SEMN_ACCDN: Day of the week or date of collision.
• GRAVITE: Severity of collision.
• CD_GENRE_ACCDN: Kind of collision.
• CD_ETAT_SURFC: Condition of rolling surface during the collision.
• MRC: Name of the regional county municipality.
• RUE_ACCDN: Street name where collision occurred.
• ACCDN_PRES_DE: Landmark near collision site.
• CD_COND_METEO: Weather conditions.
• CD_ECLRM: Illuminance on the road.
• CD_ENVRN_ACCDN: Environment of dominant activity in the area.
• VITESSE_AUTOR: Authorized speed on the road.
• LOC_LAT: Latitude (WGS84).
• LOC_LONG: Longitude (WGS84).

All the collisions are divided into five classes of severity, namely property damage, damage below reporting threshold, minor, serious, fatal.

5.2. Data Preprocessing

Preprocessing is required to get the dataset ready for analysis [55]. It involves cleaning the data and fixing the missing values and inconsistencies. Columns missing more than 50% of the values were removed as the impact on the outcome was negligible. Columns with less than 50% missing values were managed by filling the missing values with column mean for the numerical column, and for columns that represent categorical variables, they were filled with the value from the previous row. The dataset was examined for discrepancies in information formats, such as temporal attributes (DT_ACCDN and H_ACCDN) and were converted to numerical format. The necessary adjustments were made to ensure consistency.

5.3. Data Analysis

Data analysis and visualization was used to detect patterns and anomalies. The accidents were analyzed based on severity, hours of occurrence, types, hour of occurrence, weather condition, road speed limit, and location.

Severity-Based Selection: Records were visualized based on the severity level of road accidents, allowing a focused analysis on high-risk incidents. As shown in Figure 4, most of the accidents that happened from 2012 to 2021 are under the two largest segments “Damages of materials Below Reporting Threshold (39.9%)” of USD 2000 and “Damages of material Only (38.1%)”. The most concerning are the Serious (0.8%) and Fatal (0.3%) crashes that represent the smallest slices of the chart, totaling just around 1%.

Hours of the Accidents: Data was filtered for specific time frames to analyze accident trends during peak and off-peak hours. Figure 5 presents the hour of fatal and serious accidents.

• Non-serious accidents: The largest number of accidents are from afternoon to early evening from 3 p.m. to 5 p.m. This shows a link between higher traffic levels during workday commutes and midday travel. This peak is seen primarily in categories like minor accidents, property damage only, and damage below reporting threshold. Incidences of accidents decrease in the evening after 07:00 p.m., and counts are continuously lower throughout the overnight hours 0:00 a.m.–5:00 a.m.
• Serious accidents: Although their occurrence is very low, i.e., 0.8%, they also seem to increase in the same peak hours of early evening and are most likely due to increasing traffic in the peak rush hour during evenings.
• Fatal accidents: Although they occur the least, i.e., 0.1%, they are the most significant type of accidents as they might result in mortality. They do not have any pattern in daily occurrence, which may imply that they are not the result of traffic volume but some kind of human error involvement.

Overall, the graph indicates that afternoon and early evening are the most collision-prone times in Montreal, aligning with typical high-volume traffic periods. Therefore, in our simulation model, we will consider traffic volume of the peak hour period from 3:00 p.m. to 5:00 p.m.

Types of collisions: Figure 6 shows the number of collisions categorized by kind, with the height of each bar denoting the proportion of collisions (CD_GENRE_ACCDN from the dataset). Below is a summary and comparison:

• Most of the collisions occur in the vehicle category. In general, it refers to collisions (such as car-on-car crashes) that include at least two cars.
• The second most frequent kind of accident happened with “collision with pedestrian”.
• The third most frequent type of collision happened with cyclists, followed by fixed objects such as lampposts, guardrails, or other roadside constructions.
• The rest of the collisions have much lower value, so we will only focus on the first one, which is vehicle-to-vehicle collision.

Weather condition: Figure 7 shows that a significant portion of accidents occurs during clear weather, followed by the “Rain” and “Cloudy” weather conditions. This implies that mishappenings on roads are not caused solely by adverse weather conditions but also due to human error and road conditions.

Road speed Limit: Figure 8 shows the speed limit with the highest frequency of accidents on the road is 50 km/h, which is significantly more than the other categories. Most accidents occur within the limit range of 30–50 km/h.

Location: The dataset was filtered based on the most occurring street name (“RUE_ACCDN”) and the nearest intersection on that street (“ACCD_PRES_DE”). Figure 9 presents the results. It can be seen that the Sherbrooke/Papineau intersection is one of the most accident-prone areas in the Montreal municipality (MRC).

6. Case Studies

Two case studies are designed to simulate accidents in the presence of autonomous vehicles and human-driven vehicles and test how the reaction time of drivers impacts the road safety and number of accidents.

• The first case study analyzes how the increase in reaction time of HDVs will impact the number of road collisions.
• The second case study is a collision comparison of increasing penetration rate of a FR vehicle (level 3 autonomous vehicle) with an SR human-driven vehicle. It will help to analyze how AVs will behave in a mixed traffic scenario.

The simulation network creation was done by importing the geolocation net.xml from OpenStreetMap. Netedit was used to further define and fine-tune the network.

Table 1. Parameters of the road network in simulation.

Parameters Number	Parameter Name	Parameter Value
1	Road Type	Urban Intersection
2	Road length	1070.81 m
3	Maximum Speed Allowed	13.89 m/s (50 km/h)
4	Green time in one phase	42 s
5	Red time in one phase	46 s
6	Yellow time in one phase	4 s
7	Number of Vehicles	8457
8	Lane Configuration	Multi-lane

presents the details of the road network in a simulation. The screenshots of the road network created with the help of OSM web wizard and Netedit for simulation are depicted in Figure 10.

6.1. Traffic Flow Modeling in Simulation

Traffic volume data is sourced from Comptages Des Véhicules, Cyclistes Et Piétons Aux Intersections Munies De Feux De Circulation [88], for Sherbrooke/Papineau intersection on 2022-05-12 during the peak hours of 3 p.m. to 5 p.m., as shown in the origin–destination matrix (

Table 2. Model parameters and constant values in each scenario.

Model Parameter Definition	Constant Parameter in Each Scenario
Vehicle Reaction time to surroundings	actionStepLength value updating in each scenario of case study 1
Simulation step	stepLength = “1.0”
Perceived deceleration of leading vehicle (m/s2)	apparentDecel = “4.5”
Acceleration (m/s2)	accel = “2.6”
Deceleration (m/s2)	decel = “4.5”
Driver imperfection coefficient	sigma = “0.5”
Desired Time Headway (s)	tau = “1.0”
Minimum gap (m)	minGap = “2.5”
Expected multiplier for lane speed limits	speedFactor = “1.0”
Deviation of speed factor	speedDev = “0.1”
Vehicle length (m)	length = “4.5”
Maximum Velocity of vehicle (KM/H)	maxSpeed = “50”

). Each row represents the number of vehicles for each direction. The rows have the sum value of the combined traffic flow for 2 h from 3 p.m. to 5 p.m,. and the routes that are not allowed due to turn signs and traffic rules have 0 value. Each possible direction is represented by an Edge ID in the SUMO network file, and these edges are then used to create all possible routes. The same Edge IDs are mentioned under each direction in the Origin Destination matrix. For the Sherbrooke/Papineau intersection, there are a total of seven routes possible according to our simulation.

• SBLT: South-bound left turn;
• SBT: South-bound through;
• SBRT: South-bound right turn;
• EBT: East-bound through;
• EBRT: East-bound right turn;
• WBLT: West-bound left turn;
• WBT: West-bound through.

6.2. Krauss Model Parameters

SUMO uses the Krauss car following the model by default. The following model parameters were used in the Krauss model for HDV vehicle types in simulation.

Table 3. Origin destination matrix for all possible routes.

	Origin	SBLT	SBT	SBRT	EBLT	EBT	EBRT	WBLT	WBT	WBRT
		25876960#0	25876960#0	25876960#0	0	455449061#0	455449061#0	-20418390#2	-20418390#2	0
Destination
SBLT 20418390#1		79	0	0	0	0	0	0	0	0
SBT 37625473#1		0	3230	0	0	0	0	0	0	0
SBRT-455449061#3		0	0	154	0	0	0	0
EBLT		0	0	0	0	0	0	0	0	0
EBT 20418390#1		0	0	0	0	1875	0	0	0	0
EBRT 37625473#1		0	0	0	0	0	840	0	0	0
WBLT 37625473#1		0	0	0	0	0	0	840	0	0
WBT-455449061#3		0	0	0	0	0	0	0	1439	0
WBRT		0	0	0	0	0	0	0	0	0

shows the default values of parameters used for HDVs except “actionStepLength” and “maxspeed”, which are assumed per requirement.

6.3. Case Study 1

The goal of this case study is to identify if Fast Reacting (FR) vehicles enhance the road safety and reduce the number of accidents as compared to Slow Reacting (SR) vehicles. The microscopic simulation is executed on the signalized road intersection network. The HDV reaction time of “2.0” is the least possible non-default reaction time, which can be considered for HDVs, as actionStepLength should be a non-negative multiple of the simulation step length, which in our case is “1.0”. The rest of the scenarios have a gradual increase of 1 s in reaction time (actionStepLength), as mentioned in

Table 4. Default parameters for HDV of the Krauss model [53].

Model Parameter Definition	Constant Parameter in Each Scenario
Vehicle Reaction time to surroundings	actionStepLength value updating in each scenario of case study 1
Simulation step	stepLength = “1.0”
Perceived deceleration of leading vehicle (m/s2)	apparentDecel = “4.5”
Acceleration (m/s2)	accel = “2.6”
Deceleration (m/s2)	decel = “4.5”
Driver imperfection coefficient	sigma = “0.5”
Desired Time Headway (s)	tau = “1.0”
Minimum gap (m)	minGap = “2.5”
Expected multiplier for lane speed limits	speedFactor = “1.0”
Deviation of speed factor	speedDev = “0.1”
Vehicle length (m)	length = “4.5”
Maximum Velocity of vehicle (KM/H)	maxSpeed = “50”

, while keeping the other parameters as constant. In total, 8457 vehicles are loaded in simulation, which will run on pre-determined routes for all the scenarios.

6.4. Case Study 2

The goal of this case study is to find how AV’s behave as compared to human drivers on the same simulation network and how they will impact the number of road collisions. In the microscopic simulation, autonomous cars with level 3 automation are introduced.

• H0 (Null Hypothesis): “There is no significant difference in the number of collisions between HDV vehicles and AV vehicles”.
• H1 (Alternative Hypothesis): “Significant reduction in collisions with AV vehicles”. Cars with level 3 automation are introduced.

To test the hypothesis, “tau”, which represents the minimal headway, “sigma”, which accounts for driver flaws, and “actionStepLength”, which refers to vehicle reaction time to its surroundings, are used. In this research, the HDVs are assumed to have a slower reaction time “2.0” as compared to AVs; this is to signify human reaction times are inherently variable and influenced by factors like fatigue, distraction, and impairment. Studies consistently show that even under optimal conditions, human reaction times are significantly slower than the processing speeds of AV sensors and computers. The HDV reaction time of “2.0” is the least possible non-default reaction time, which we can consider for HDVs, as actionStepLength must be a non-negative multiple of the simulation step-length, which in our case is “1.0”. The AV is assumed to have the shorter perception-reaction time “1.0”. The value of “tau” and “maxspeed” is kept the same for both HDV and AV, while the rest of the parameters are derived using the literature and are presented in

Table 5. Scenarios based on different reaction times.

Scenarios (Reaction Time)	Vehicle Type	Variable Parameter
Scenario 1	Human Driver Vehicle	actionStepLength = “2.0”
Scenario 2	Human Driver Vehicle	actionStepLength = “3.0”
Scenario 3	Human Driver Vehicle	actionStepLength = “4.0”
Scenario 4	Human Driver Vehicle	actionStepLength = “5.0”
Scenario 5	Human Driver Vehicle	actionStepLength = “6.0”

.

Simulation Scenarios in Case Study 2

Autonomous cars with level 3 automation are introduced gradually, as mentioned in

Table 6. Krauss CF model parameters for HDV and AV [54].

Parameters	HDV	AV (Level 3)
Acceleration (m/s2)	accel = “2.6”	accel = “3.6”
Decceleration (m/s2)	decel = “4.5”	decel = “4.5”
Driver imperfection coefficient	sigma = “0.5”	sigma = “0.2”
Desired Time Headway (s)	tau = “1.0”	tau = “1.0”
Minimum gap (m)	minGap = “2.5”	minGap = “1.25”
Simulation step	stepLength = “1.0”	stepLength = “1.0”
Vehicle Reaction time to surroundings	actionStepLength = “2.0”	actionStepLength = “1.0”
Maximum Velocity (KM/H)	maxSpeed = “50”	maxSpeed = “50”

. The goal of this microscopic simulation is to find how AVs behave as compared to human drivers on the same simulation network and how will they impact the number of road collisions.

Table 7. Scenarios based on AV penetration rate.

Scenarios	AV Penetration Rate	Total Number of Vehicles
Scenario 1	0%
Scenario 2	25%
Scenario 3	50%	8457
Scenario 4	75%
Scenario 5	100%

presents the scenarios and AV penetration rates.

7. Results

7.1. Results of Case Study 1

Table 8. Results of case study 1; each sprint represents execution run with different seed values.

Scenario	Sprint	Sprint	Sprint	Sprint	Sprint	Sprint	Sprint	Sprint	Sprint	Sprint
	1	2	3	4	5	6	7	8	9	10
Scenario 1	5436	5425	5393	5342	5159	5355	5409	5516	5674	5485
Reaction
time 2
Scenario 2	6490	6529	6530	6531	6701	6623	6676	6654	6603	6838
Reaction
time 3
Scenario 3	7110	7174	7104	7081	7130	7155	7102	7179	7261	7218
Reaction
time 4
Scenario 4	7651	7670	7815	7599	7698	7679	7588	7663	7659	7679
Reaction
time 5
Scenario 5	8401	8493	8389	8447	8398	8445	8489	8402	8421	8456
Reaction
time 6

shows the number of road collisions for each scenario where reaction time is increased by 1 s. Running the simulation using the runSeeds.py for 10 times with different seeds allowed us to observe how sensitive the simulation results are to randomness, which is important due to its stochastic nature. Each sprint represents an iteration of the simulation run.

The heatmap in Figure 11 shows the distribution of the collision types between the traffic vehicles. There are two types of collisions observed in our simulation:

• Rear collision: The rear-end collision signifies when a follower vehicle crashes into the leader vehicle; therefore, in that case, the leading vehicle is the victim.
• Junction collision: This means a collision between vehicles at a junction or intersection. In this case, the collider and victim are assigned arbitrarily by SUMO.

Rear collisions represent more than 99% of collisions in all the scenarios, and their number keeps on increasing as the reaction times of HDV are increasing. Junction collisions are negligible in all the scenarios, which may be due to the signalized intersection that guides drivers to follow a sequence and hence prevent the collisions.

7.2. Results of Case Study 2

Table 9. Results of the microscopic simulation for case study 2.

AV Penetration Rate	Sprint 1	Sprint 2	Sprint 3	Sprint 4	Sprint 5	Sprint 6	Sprint 7	Sprint 8	Sprint 9	Sprint 10
0%	5436	5425	5393	5342	5159	5355	5409	5516	5674	5485
25%	4283	4040	4049	3950	4163	3982	4058	4012	4445	4122
50%	2775	2911	2776	2816	2813	2875	2908	2988	2887	2797
75%	1525	1504	1569	1500	1432	1463	1565	1423	1502	1412
100%	0	0	0	0	0	0	0	0	0	0

presents the impact of autonomous vehicle (AV) penetration rate on the number of collisions across multiple simulation sprints, each sprint represents an iteration of the simulation run. To handle the stochastic nature of the simulation, 10 iterations with different seed values were performed for each scenario to obtain reliable outputs.

Autonomous vehicle penetration rate reduces the number of road collisions significantly from 0% PR to 100% PR. As seen in

Table 10. Distribution of collision results between each vehicle type.

AV Penetration Rate	AV-AV	HDV-AV	HDV-HDV	Mean of Total Collisions
0% (base case)	0	0	5419.4	5419.40
25%	0.2	2249.5	1860.7	4110.40
50%	0.2	2164.20	690.2	2854.60
75%	1.7	1321.5	166.3	1489.50
100%	0	0	0	0

, the mean number of collisions decreases from 5419.40 to 0. The mean of total collisions shows the overall average number of collisions over the 10 sprints, which are decreasing with the rise in AV penetration rate.

Figure 12 represents the distribution of collision types between the traffic vehicles. There are two types of collisions observed in our simulation.

The t-statistics-based hypothesis test between scenario 1, which has 100% HDV traffic, and scenario 5, which has 100% AV traffic, shows significant statistical difference with a t-value of 129.5425 and degree of freedom equal to 18. Hence, this results in a p-value of less than 0.05, which means the rejection of the null hypothesis (H0) and proves that the alternate hypothesis (H1) is correct, i.e., “Significant reduction in number of collisions with AVs”.

7.3. Discussion

The results of case study 1 suggest that a longer reaction time while driving HDVs (e.g., due to human factors like distraction, fatigue, or slower processing) leads to a higher number of road collisions. This finding suggests important implications for efforts to improve road safety through technical breakthroughs and interventions that reduce human error. Autonomous vehicles could be a potential solution as they are designed to have much faster and more consistent reaction times than human drivers. All the results and graphs of case study 1 imply that shorter reaction times minimize the number of collisions, whereas longer reaction times considerably increase accident risk, which may be because the slower reaction time increases the perception time of drivers. Hence, the response to the changes in the driver’s environment is delayed, and this discovery could be critical when comparing human drivers to autonomous vehicles, as AVs often have substantially lower reaction times.

Case study 2 results suggest that AVs can significantly enhance the road safety and decrease the road conflicts. In the baseline scenario, 64.08% of road accidents were happening but with the introduction of 25% AVs, a reduction of almost 24.15% was observed and the number of collisions went down to 48.60%. The number of collisions went down to 33.75%, 17.61%, and 0% as the AV penetration reached to 50%, 75% and 100%. This may be because of the fact that they eliminate the accidents taking place due to human error such as fatigue, rash driving, slow reaction time (distraction) and impairment. On the other hand, AVs have ADAS, faster reaction time, high safety equipment, and advanced sensors. The studies by Morando et al. [89], Papadoulis et al. [90] and Karbasi et al. [81] show similar simulation results that AVs reduce the road conflict to a significant percentage and eliminate the errors caused by human drivers due to distraction, fatigue, slow reaction or impairment.

8. Conclusions and Future Research

In this paper, we investigate the impact of autonomous vehicles on road safety performance. Firstly, a detailed analysis of the road accidents in Montreal from 2012 to 2021 is presented. It sheds light on the various insights of accidents such as collision types, days of the week, speed limit, weather, and road conditions. Critical accident-prone zones and peak collision times offer a localized insights concerning Montreal. The data implied that one of the reasons behind most of the accidents is some kind of human error. Two case studies are performed to measure human-driven vehicle and autonomous vehicle performance.

The results of case study 1 represent a positive correlation between the reaction time and number of collisions observed; even a small increase in the reaction time of drivers resulted in a high number of collisions. There were approximately 3000 more accidents when drivers’ reaction time was increased from 2 to 6 s. Case study 2 shows AVs and HDVs behavior in mixed traffic scenarios and how they impact the number of collisions. The results show a significant reduction in road accidents as the autonomous vehicle PR from 0% to 100% was reached.

This study highly recommends a hybrid solution of human driving with Advanced Driver Assistance System (ADAS) technology as a valuable and fast-track solution with our current infrastructure as it provides features such as emergency braking and lane-keeping assist. Policy makers should initiate more pilot programs to assess the real-world implications of level 5 AVs adoption in lightweight vehicles and transportation sections and carefully assess their interactions with existing HDVs.

This study focuses on the urban environment of Montreal, which has unique characteristics such as traffic density, road intersections, speed limits, accident-prone locations and timings. While the findings provide valuable insights into accident patterns, the role of reaction time and autonomous vehicles in improving safety, caution should be taken when applying these results directly to other geographical areas.

However, the underlying principles such as the positive impact of faster reaction time and the behavior of AVs in mixed traffic are likely applicable to other urban environments with similar traffic complexities. Future research could extend this analysis to different geographical settings to strengthen the external validity of the conclusions.

8.1. Limitations

Although SUMO is a robust platform, real-world driving is a complex behavior, so aspects of simulating human actions, vehicle interactions, maneuvering behavior and driver conducts are limited. This research does not study the risks that come with autonomous vehicles, such as cyber-attacks, user privacy, and onboard system failure. The simulations cannot incorporate all the real-world parameters of humans and roads. This study only focuses on rear-end and junction collisions. This traffic in this study consists of lightweight vehicles only and focuses on the number of road accidents.

8.2. Future Research

In this paper, we used a microscopic simulation in the evaluation of the safety consequences of AVs. However, the proposed approach has limitations that must be addressed in future research.

The relevant data to simulate the behavior of autonomous vehicles was limited as the Quebec government currently only allows level 3 AVs on the road. Further studies should be done to explore the same model with a higher automation level of AVs.

This study only focused on the number of road accidents with different reaction time and different vehicle types. In the future, the road efficiency of vehicles with different levels of automation can be studied.

Our study used a single model. In the future, more models such as SSAM and CACC could be incorporated for a better understanding of AVs. New simulation case studies can be created based on the comparison of different automation levels of AVs and performing sensitivity analyses of their model parameters such as minimum gap, acceleration, deceleration, and tau.

More research should be conducted on the risks posed by the rising penetration of AVs on roads, specifically in heterogeneous traffic conditions.