2.1

Data preparation

In this paper, 401 accident reports from 2019 to 2023 were collected from the crash data source of autonomous vehicles published on the official website of California Motor Vehicle Administration^[12]. The data cleaning rules are as follows: (1) Whether there is any major omission in the accident record. (2) Whether the autonomous vehicle was in automatic driving mode at the time of the accident. Finally, 197 valid autonomous vehicle accident records were obtained.

A preliminary analysis of the accident data found that the manufacturers of autonomous vehicles in the accident report included: Apple, Argo AI, GM Cruise, Lyft, Mercedes-Benz, Pony, Waymo, Weride, Zoox. 21.8% of the accidents resulted in personal injury, and 78.1% of the accidents had no or only property damage. 64.5% of the collisions occurred at the rear of the vehicle, 21.5% at the rear of the vehicle, 6% at the middle of the vehicle, 15.5% at the front and 9% at the head of the vehicle. It can be seen from the damage of the car body in Figure 1 that the rear-end collision accident dominates.

Figure 1.

Vehicle damage location distribution

2.2

Classification of dangerous scenarios

NHTSA 37 Preliminary Crash Scenarios is a set of vehicle crash test standards developed by the National Highway Traffic Safety Administration (NHTSA) to evaluate the safety performance of vehicles^[13]. The test standard divides the common road scenarios into 37 categories, and tests the collisions and the dynamic characteristics of the vehicle during the collision under each category of scenarios. In recent years, with the development of autonomous driving technology, NHTSA 37 pre-crash scenario classification has also been applied to evaluate the safety performance of autonomous vehicles. This is because autonomous vehicles need to make the right decisions and ensure the safety of surrounding vehicles in a variety of road and traffic conditions, so the testing of NHTSA 37 pre-crash scenario classification can provide guidance and support for the development of autonomous vehicles.

To identify typical hazard scenarios for autonomous vehicles, 197 crash records were matched to the NHTSA Class 37 pre-crash scenario classification. The matching is based on the structured and unstructured information in the crash record, including the text description of the crash, the vehicle state at the time of the accident, the vehicle dynamics before the collision and the collision type. The classification results are shown in Table 1. It can be seen that the top two occurrence frequencies of typical dangerous scenarios of automatic driving are Lead Vehicle Stopped and Following Vehicle Making a Maneuver respectively.

Table 1.

Classification results of dangerous scenarios.

2.3

Extraction of Dangerous Scene Elements

In order to further build the simulation model of typical dangerous scenes of autonomous driving, it is also necessary to analyze the crash-related factors and determine the typical scene elements that have significant influence on the crash. In this study, random forest algorithm is selected to extract typical scene elements by regression with crash-related influencing factors as independent variables and crash casualties as dependent variables. Random Forest is an ensemble learning-based algorithm that performs regression tasks by building multiple decision trees and integrating their predictions^[14]. In a random forest, each decision tree is independent and trained on randomly selected subsamples, which effectively reduces the risk of overfitting. The results are shown in Table 2.

Table 2.

Importance of random forest regression features.

According to the results of random forest regression analysis, the typical scenario factors affecting the severity of accidents that the importance is more 5.0% are: pre-collision vehicle movement, location of vehicle damage, extent of vehicle damage, manufacturer, type of collision, weather and lighting conditions.

3.1

Dangerous scenario simulation Settings

In the scene of Lead Vehicle Stopped, a simulation control flow is set as follows: Both the front and rear vehicles drive at the intersection with the given accelerator. When the two vehicles approach, the signal light turns yellow, the front vehicle (self-driving vehicle) brakes immediately, and the rear vehicle (manual driving vehicle) avoids. The CARLA simulation scenario is shown in Figure 2.

Figure 2.

Simulation scenario of Lead Vehicle Stopped.

In the scene of Following Vehicle Making a Maneuver, the simulation control flow is set as follows: Both the front and rear vehicles drive at the given accelerator at the intersection position. When the two vehicles approach, the signal light turns yellow, the front vehicle (autonomous vehicle) still drives at the original speed, and the rear vehicle (manual driving vehicle) turns right. The CARLA simulation scenario is shown in Figure 3.

Figure 3.

Simulation scenario of Following Vehicle Making a Maneuver.

3.2

Simulation output design

In order to realize the subsequent extraction of safety indicators and risk analysis, two output modules are set in this study.

Collision detection module: The vehicle position in CARLA is represented by the coordinates in the three directions of the vehicle chassis. Therefore, the check_collision function is defined in this study, and the Euclidean distance is used to detect whether the vehicle collides. The Euclidean distance represents the true distance between two points in a dimensional space and is calculated as follows:

Where n is the coordinate dimension; and x_i are the coordinate value.

Vehicle distance and speed output module: In order to extract safety indicators, the speed and position information of the vehicle will be output during each execution cycle 0.1s.

4.1

Safety assessment indicators

As an open source simulator dedicated to autonomous driving research, CARLA can provide speed and distance information accurate to 0.1 seconds, which is difficult to obtain by traditional conflict research methods. Since speed and distance are both important to measure traffic safety, and in the study of traffic accidents relative speed and distance can reflect the risk and severity of accidents more intuitively. Thus this study selects the calculated relative speed and diatance of the front and rear vehicles as the safety assessment index.

Furthermore, in the research of traffic conflict technology, TTC has been proved to be an effective means to measure the severity of traffic conflict and to distinguish critical behavior from normal behavior. TTC is one of the indicators for evaluating the driving safety of vehicles. Under the same driving conditions, the larger the value of TTC, the safer the vehicle is ^[15]. Therefore, TTC is also selected as the safety assessment indicator. The calculation formular is:

Where Δd is the relative distance between the front and rear vehicles; Δv is the relative speed of the front and rear vehicles. A lower TTC value indicates a more dangerous situation.

4.2

Risk analysis of the scene ‘Lead Vehicle Stopped’

In the simulation experiment of the scene ‘Lead Vehicle Stopped’, according to the yellow light duration and the urban road speed limit of 30km/h-60km/h, two kinds of following distance of 15m and 20m are selected to carry out the experiment. By changing the different acceleration distance, the vehicle has different initial velocity when reaching the specified following distance. After calculation, the simulation results of the braking scenario of the preceding vehicle are shown in Figure 4 and 5. Different curves represent different initial speeds at the specified following distance in the current experiment.

Figure 4.

Simulation Results of the scene ‘Lead Vehicle Stopped’ (following distance of 15m).

Figure 5.

Simulation Results of the scene ‘Lead Vehicle Stopped’ (following distance of 20m).

According to the experimental results, in the ‘Lead Vehicle Stopped’ scenario, the change trend of relative speed, relative distance and TTC of the two vehicles is roughly the same when the following distance is 15m and 20m, but the vehicle collision risk when the following distance is 15m is significantly higher than that when the following distance is 20m.

When the following distance is 15m, the relative distance between the two vehicles decreases faster with the increase of the initial speed of the following vehicle, and the TTC value is almost always within the range of less than 1.5. When the initial speed of the following vehicle is less than 16.5m/s, the following vehicle can successfully avoid collision with the preceding vehicle by means of lane change and sudden braking. When the initial speed of the following vehicle is greater than 16.5m/s, the TTC value will always be within the range of 0.6, and will continue to decrease with the increase of time. At this time, the following vehicle cannot avoid collision with the preceding vehicle regardless of the way of lane change or sudden braking. It should be noted that, since the default vehicle length in this experiment is 4.5m, when the initial speed of the following vehicle is between 13.1m/s and 16.5m/s, even if the following vehicle can take measures to change lanes or brake in time to avoid collision with the preceding vehicle, the minimum distance between the two workshops is still less than 2m, so collision is still likely to occur according to the difference in the length of the vehicle body in reality.

At a following distance of 20m, the TTC value is almost always in the range of 0.5 to 2.0, and the collision risk is significantly lower than at a following distance of 15m. When the initial speed of the following vehicle is in the range of 8.2m/s~16.3m/s, the following vehicle can always avoid collision with the preceding vehicle by steering and braking. The minimum distance between the two workshops is always greater than 4m, that is, about one vehicle length, which is lower than that when the following distance is 15m.

To sum up, at the end of the simulation test of the ‘Lead Vehicle Stopped’ scenario, the distance between the two vehicles is smaller when the initial following distance is 15m than that when the initial following distance is 20m, and the risk of accident is greater. It can be seen that when the following distance is 20m, the following vehicle can have longer reaction and operation time, and can better respond to the sudden behavior of the autonomous vehicle in front.

4.3

Risk analysis in rear turning scenario

In the simulation experiment of Following Vehicle Making a Maneuver scene, the following distance of 15m and 20m is also selected to carry out the experiment. By changing the different acceleration distance, the vehicle has different initial velocity when it reaches the specified following distance. After calculation, the simulation results of the braking scenario of the preceding vehicle are shown in Figure 6 and 7.

Figure 6.

Simulation Results of the scene ‘Following Vehicle Making a Maneuver’ (following distance of 15m).

Figure 7.

Simulation Results of the scene ‘Following Vehicle Making a Maneuver’ (following distance of 20m).

According to the experimental results, in the scene of ‘Following Vehicle Making a Maneuver’, the change trend of relative speed, relative distance and TTC of the two vehicles is roughly the same when the following distance is 15m and 20m, but the collision risk of the vehicle with the following distance of 15m is still significantly higher than that of the vehicle with the following distance of 20m.

When the following distance is 15m, the relative distance between the two vehicles decreases faster with the increase of the initial speed of the following vehicle, and the TTC value is almost always within the range of less than 2.0. When the initial speed of the following vehicle is less than 12.5m/s, the following vehicle can turn successfully and does not collide with the preceding vehicle. However, when the initial speed of the following vehicle is greater than 12.5m/s, the TTC value will always be within the range of 0.8, and the probability that the following vehicle can successfully turn is very small, and effective avoidance cannot be achieved within 0.8 seconds, so it is almost inevitable to collide with the preceding vehicle.

At a following distance of 20m, the TTC value is almost always in the range above 1.0, and the collision risk is significantly lower than at a following distance of 15m. When the initial speed of the following vehicle is above 12.5m/s under the two following distances, the vehicle with the following distance of 20m can also successfully avoid collision with the preceding vehicle by braking, and the minimum distance between the preceding and the following vehicles is always greater than 5m, so the collision risk is low.

Therefore, in the Following Vehicle Making a Maneuver scenario, the rear vehicle with an initial following distance of 20m can better control the speed and distance between vehicles when steering, and the reaction time for braking action is longer, and the risk of accident is significantly lower than that of the front vehicle with an initial following distance of 15m.