Key points of investigation and analysis on traffic accidents involving intelligent vehicles

Abstract

Intellectualization is regarded as the future mainstream development trend of the automobile industry. The automation level of intelligent vehicles is relatively low so far, and the road traffic system will be in a mixed state of non-autonomous vehicles and vehicles with different levels of automation for a long time. Therefore, the road traffic system will be more complex with more diverse accidents. This paper analysed the characteristics and causal factors of intelligent vehicle accidents. Based on the problems existing in investigation, analysis and liability identification of intelligent vehicle accident, the study proposed a preliminary accident investigation framework and method, summarized the key points of accident analysis from the perspectives of technical defects, information security and passive safety, and specified the liability subjects for intelligent vehicle accidents and their corresponding legal liability. The results from this study contributed to the development of intelligent vehicle accident investigation and disposal, and provided the reference for the improvement of vehicle safety and accident prevention.

1. Introduction

With the development of computer and communication technology, great achievements have been made in the field of intelligent vehicles. Intelligent vehicles, which fundamentally change the traditional ‘human-vehicle-road’ closed-loop control method, greatly improve road usage efficiency and road traffic safety by eliminating unstable drivers from the closed-loop system. As the development trend of future automotive technology, intelligent vehicles have shown broad application prospects and obtained huge economic benefits. At present, these vehicles with various technical conditions and automatic levels are in a state of co-existence, which makes the road traffic system more complex and results in more diverse road traffic accidents. Furthermore, laws and regulations on intelligent vehicles in various countries are still incomplete, which also leads to a variety of social problems.

Intelligent vehicles, which perceive the surrounding environment based on on-board sensors and make decisions based on artificial intelligence technology, can autonomously navigate and reach a destination without manual operation. Google took the lead in researching intelligent vehicles and conducted a series of road tests in California in 2010. The first intelligent vehicle accident, in which a Google intelligent vehicle scratched a bus during merging due to misjudgment of the other driver's intention, occurred in february 2016 [1]. In 2015, Tesla developed the auto pilot assisted driving system, including automatic lane-keeping, automatic lane-changing and automatic parking. In addition, an intelligent vehicle accident, in which a Tesla vehicle failed to recognize a white trailer and collided with it when passing an intersection, occurred in May 2016 [1]. This type of collision indicates that the birth of intelligent vehicles will lead to new traffic conflicts.

Compared with traditional vehicle accidents, intelligent vehicle accidents are more complex. In the future road traffic system, new traffic phenomena, such as the co-existence of various autonomous and human-driven vehicles, man–machine co-driving, information security of the intelligent traffic system, occasional mechanical failure, environmental and weather influence, may lead to intelligent vehicle accidents. Therefore, it is necessary to create intelligent vehicle accident investigation methods, and to understand the characteristics and mechanisms of intelligent vehicle accidents to better solve the current technical defects of intelligent vehicles, and provide solid theory and data support for traffic safety management, accident prevention and rescue, and formulation of relevant laws and regulations. This paper involves the following points, (1) analysing accident characteristics based on the reported intelligent vehicle accident data; (2) constructing a new intelligent vehicle accident investigation system based on existing experience and traditional methods; (3) summarizing the key content of intelligent vehicle accident analysis from the perspectives of technical defects, information safety, passive safety, etc.; and (4) putting forward relevant recommendations for the determination of liability for intelligent vehicle accidents.

2. Analysis on characteristics of intelligent vehicle accidents

2.1 Intelligent vehicle

Intelligent vehicles can perform autonomous driving based on a series of on-board intelligent hardware. Automatic driving, which blends cognitive science, artificial intelligence, robotics and vehicle engineering, is regarded as an important development direction of modern science and technology. In 2014, the Society of Automotive Engineers (SAE) classified automatic vehicles into five levels according to the degree of automation, as shown in Table 1 [2]. This classification contributes to analysing accident causation and determining who was responsible for the accident before intelligent vehicles are fully automated.

The classification indicates that the development of automatic driving cannot be achieved overnight. Therefore, for some time to come, the road traffic system will be in the state where vehicles with various technical conditions and automated levels are mixed. The road traffic system will be more complex, and traffic accidents will also be more diverse. The reported intelligent vehicle accidents are summarized in Table 2.

2.2 Contributing factors of intelligent vehicle accidents

Limited by the current low automation level, intelligent vehicles inevitably conflict with the existing man–vehicle–road system. As far as accident causation is concerned, intelligent vehicle accidents have three major characteristics based on the idea of man–machine–environment system engineering.

2.2.1 Intelligent vehicle accidents are mainly caused by machine factors

Based on artificial intelligence technology, intelligent vehicles complete autonomous decision-making and operation depending on information perception modules (radars, cameras, V2X equipment, etc.), information processing modules (deep learning, etc.) and execution modules. However, in the development process of smart vehicles, the technology has been gradually perfected. It is inevitable that there will be technical limitations to a certain extent, which may easily cause collision accidents due to technical problems. For example, due to the shortcomings of the current intelligent recognition technology, a Tesla vehicle collided with a white trailer in Florida in May 2016; due to the limitation of intelligent decision-making, a Google intelligent vehicle scratched a bus during merging after misjudging the other driver's intention in California in February 2016 [1]. In particular, intelligent vehicles are not friendly for vulnerable road users (VRUs) due to the unreliable interactions between VRUs and intelligent vehicles. At present, intelligent vehicles still have difficulty in detecting and predicting VRUs due to more abundant postures, and more causal motions, etc. (e.g. Google-cyclist in Austin in 2015, Uber nearside hooks in San Francisco, Tesla's autopilot's difficulty in detecting cyclists), which puts higher requirements on computer image recognition and prediction algorithm. In addition, mode confusion, which is an unexpected system response during an attempt to activate or take over the automatic driving, may confuse the driver about the current state of the intelligent vehicle, and may also lead to an intelligent vehicle accident. Furthermore, information security has become extremely important in the era of intelligent networking. The safe information transmission and reliable system have important influence on the safety of the intelligent transportation system. Different degrees of system disorders resulting from possible attacks on computer information and communication systems from the external (such as hackers) or unexpected situations, such as the small probability of system errors, may cause errors in the interaction among human, vehicle and road [1,3].

2.2.2 Human factors are mainly for operating the intelligent vehicle without following the instructions

In the transition stage from non-autonomous driving to fully automatic driving, existing intelligent vehicles cannot do without the assistance of human drivers. The assistance of human drivers can make up for the shortcomings of current automatic driving technology. A reasonable time interval should be reserved before switching from human operation to machine operation, otherwise it is easy to result in collision accidents when human drivers are not involved in time or even at all. For example, the accident of a Tesla, colliding with a white trailer in Florida in May 2016, involved a long period of distraction (at least 7 seconds). The NHTSA Office of Failure Investigation found there was less time for systems/drivers to detect, react and take action (less than 3 seconds) in most accidents [4]. As a result, the problem of operation switching between human and machine has become a research focus in the field of intelligent vehicles. Some studies showed that dedicated drivers have a better understanding of the surrounding environment in most of these situations. Especially with rich experience, they can make accurate predictions about the behaviour of other drivers [5].

2.2.3 Environmental factors are mainly for restricting the usage of intelligent systems

In traditional road traffic accidents, environmental factors mainly affect road users' information acquisition and vehicle operation stability. Intelligent vehicles are also unable to get rid of the influence of environmental factors. For example, road reflection characteristics change due to snow or ice, which affects the performance of LiDAR (light detection and ranging), and further affects the construction of three-dimensional maps, thereby leading to the problem that intelligent vehicles cannot follow the scheduled routes. With visual sensors recognizing information (such as lane lines and road signs), intelligent vehicles can avoid obstacles, and drive in accordance with laws and regulations. In heavy snow, however, lane lines, road signs and objects on both sides of the road (such as vehicles and buildings) will be difficult to identify because they are partially covered by snow. In addition, intelligent vehicles can only respond to specific traffic scenarios due to their low automation levels. For example, Tesla's auto-steer system is only used on highways with clear central dividing lines and lane markings. The human driver is responsible for determining whether the current traffic environment is suitable for activating the system, while the system only provides real-time warnings based on the current vehicle speed, road type and traffic volume. Therefore, an intelligent vehicle collision may be caused if the human driver ignores the working conditions of the automatic driving system.

2.3 Accident information

As far as accident investigation is concerned, a large amount of precollision information can be recorded in intelligent vehicle accidents due to the use of cameras, radars and smart systems, specifically as follows.

1. Driving behaviours before the collision. It is recorded in detail whether the driver correctly operates the intelligent vehicle according to the instructions, and the series of driving operations after taking over the intelligent vehicle.

2. Motion states of the accident participants before the collision. The motion states of each accident participant, such as movement trajectory, acceleration/deceleration time and intensity, turning time and degree, etc., are recorded in detail.

3. Visual obstacles existing in the accident. Visual obstacles are the main cause of traffic accidents. In traditional vehicle accidents, temporary visual obstacles (such as moving vehicles, and vehicles parked temporarily on the roadside) are often ignored due to the lack of information before the collision.

4. Surrounding traffic environment before the collision. Traffic environment information, such as weather conditions, lighting conditions and traffic light conditions, are recorded in detail.

Traditional vehicle accident identification mainly focuses on the consequences of the collision, such as damage to the car and personal injury. The precollision information is obtained by reasonable calculation based on the collision consequences (such as final position, vehicle damage, road traces and scattered objects). Hence, the identification results often deviate from the actual situation. However, in intelligent vehicle accidents, a large amount of precollision information is recorded, which provides accurate and reliable data support for accident identification. These information contributes to helping the appraiser to obtain comprehensive accident causation, and helping traffic police to more reasonably determine the accident liability that should be borne by the smart vehicle, driver and manufacturer.

3. Key issues of the intelligent vehicle accident investigation

Intelligent vehicles inevitably conflict with the existing human–vehicle–road system due to their low automation levels. An intelligent vehicle accident still occurs under the comprehensive influence of human, mechanical and environmental factors. With the development of vehicle intelligence and automation, the leading accident causation changes from human factors to mechanical factors. Thus, the focus of intelligent vehicle accident investigation needs to be redefined. Combining the traditional accident investigation method and the intelligent vehicle accident characteristics [6–10], a novel accident investigation framework is established in this study, which summarizes the issues that one should account for to investigate the intelligent vehicle accident causation. The framework provides theoretical guidance for future intelligent vehicle accident investigation.

3.1 Survey contents

Traffic accident investigation is a multi-disciplinary work, which collects a large number of accident-related variables from various levels of the road traffic system. In-depth traffic accident data contribute to improving road design, vehicle safety, medical services and traffic management.

The traditional road traffic system is a complex, dynamic system composed of human, vehicle, road and environment elements, in which each element interacts. With the development of intelligence and network technologies, vehicle intelligent systems play an increasingly important role in decision-making, and intelligent network-connected systems also frequently participate in information exchange. Therefore, the future road traffic system will be more complex, as shown in Fig. 1.

Haddon theory is often used to analyse the accident causation and the corresponding preventive countermeasures, which indicates that road traffic accidents are caused by the problems occurring in each subsystem or the interaction imbalance between each other. According to the Haddon analysis model, the entire traffic accident process is divided into three stages in the road traffic accident investigation: precollision, collision and postcollision [11].

Intelligent vehicle accident investigation is based on, but different from, traditional vehicle accident investigation. It not only focuses on human drivers, but also on intelligent vehicles [12]. Based on the traditional system, the intelligent vehicle accident investigation framework is shown in Fig. 2.

The investigation content is explained by different stages of the accident process. The precollision stage is divided into two intervals, (1) the interval between normal driving and danger detection, which mainly includes the basic situation and driving status of human drivers (such as age, gender, emotions, drunk driving, distraction and fatigue), and road control conditions (such as road type, traffic control, environmental conditions); (2) the interval between danger detection and accident occurrence, which mainly includes the judgment of human drivers and smart vehicles (such as estimating the behaviour of other traffic participants), the decision-making of human drivers and smart vehicles (whether the human driver decides to intervene in driving, how to brake and steer, etc.), and the situation of operation switching between human and machine. The collision phase refers to the time interval from the collision occurrence to the final stop, which mainly includes initial collision parameters (such as collision speed and contact position between participants), decision-making of human drivers and intelligent vehicles (such as braking and steering), human impact responses (such as the movement of pedestrians during a collision), and mechanical responses of intelligent vehicles (whether the steering/brake system responds accurately, etc.). The post-collision stage refers to the post-processing stage of a traffic accident, which mainly includes the accident consequences (such as the final locations of participants, vehicle damage information, personal injury information, emergency rescue and medical conditions).

In-depth accident investigation helps researchers to establish more scientific and complete ‘4E’ (engineering, education, enforcement and emergency) traffic safety intervention and improvement measures. Engineering refers to accident prevention and improvement based on engineering design methods. Education mainly refers to driving skills and traffic safety awareness training in schools and society. Law enforcement is the supervision and management of traffic behaviours by traffic management departments in accordance with relevant laws and regulations. First aid emergency includes ambulance transportation services and emergency medical treatment.

3.2 Investigation methods

Different from the traditional accident investigation, the intelligent vehicle accident investigation pays more attention to the point whether the intelligent vehicle can operate correctly before the collision, and whether human–computer interaction can be carried out normally. Unmanned aerial photography and accident video are essential for obtaining the running trajectories of intelligent vehicles before the collision and risk scenarios. The event data recorder (EDR) contributes to obtaining the working status of the intelligent vehicle during the collision. The rich investigation methods will help investigators obtain comprehensive causes of intelligent vehicle accident, which contributes to accurately remedying intelligent vehicle defects and completing intelligent vehicle accident liability confirmation. Under the current technical level, intelligent vehicles have more stringent requirements for driving scenarios. China has a vast territory and uneven urban development. Therefore, the intelligent vehicle accident investigation has a regional nature, and can accurately obtain the performances of intelligent vehicles in different risk scenarios. Based on the experience of traditional accident investigation, the method of intelligent vehicle accident investigation can be divided into four steps.

1. Data basis. Researchers remotely assist traffic police to complete the data collection at the accident site. Based on the traditional accident scene investigation, the intelligent vehicle accident investigation focuses on the driving psychology/behaviour of human driver and the working status of intelligent vehicle before the collision.

2. Data supplement. Researchers return to the accident site to perform an in-depth survey with professional equipment. The main work includes obtaining detailed environmental information of the accident site by drone aerial photography, as well as obtaining accurate working conditions of intelligent vehicles and measuring vehicle damage by extracting the data from event data recorder.

3. Data processing. Researchers visualize and digitize the accident information based on the survey results, such as drawing an in-proportion sketch of the accident.

4. Data management. Researchers input the visualized and digitized accident information into a special database for accident analysis, vehicle development and traffic management.

4. Key issues of intelligent vehicle accident analysis

Accident analysis is an important technical mean to improve vehicle technical defects and to enhance vehicle safety performance. Intelligent vehicles complete the driving by intelligent systems rather than human drivers. Drivers and passengers who are free from driving tasks can optionally adjust their postures to get greater comfort. It has become the focus of intelligent vehicle accident analysis to fully understand technical defects, information security problems and new passive safety problems. Intelligent vehicle accident analysis mainly includes the following.

1. Accident scenarios. These are the basis for the research and the development of intelligent vehicles, and are the specific units for examining the safety of intelligent vehicles. Driving scenarios with accident risk can be obtained based on in-depth learning from real accident scenarios. With the help of accelerated simulation methods and high-precision trajectory data, the test investment will fall sharply. Based on artificial intelligence technology, key scenario database assisting the intelligent vehicle to make decisions can be obtained [13, 14]. A new vehicle intelligent system has to be validated by real vehicle tests before it hits the market. The representative test scenarios, which indicate the basic requirements that must be met for an intelligent system to be applied to the vehicle, can be determined by accident scenario analysis. For the autonomous emergency braking (AEB) test in 2021 C-NCAP, the longitudinal pedestrian test scenario is added, in which the pedestrian lateral position is divided into two cases, namely 50% and 25% offset. There are also requirements for light in the AEB–Pedestrian test, which is divided into two categories, day and night. Due to the growing number of e-bike accidents in China, AEB–E-bike test scenarios have also been added.

2. Accident causation. The intelligent vehicle accident causation analysis shifts focuses from human errors to system function failures. A comprehensive analysis of the failure causes of vehicle intelligent systems should be conducted from the perspectives of information perception, information processing, decision-making and execution, such as system function failures caused by technical limitations, computer information and communication system failures caused by external (hacker) attacks and mechanical failures [15, 16].

3. Kinematic responses. Traditional vehicle accident analysis focuses on the accident results after collisions, such as impact dynamic responses and injuries. The kinematic responses, before the collision directly show whether the driving decision made by the intelligent vehicle system is reasonable. Therefore, the kinematic responses before the collisions (such as the movement trajectory, the type and response time of collision avoidance measures taken by the intelligent vehicle) should be paid enough attention in the intelligent vehicle accident analysis [17].

4. Injury biomechanics. To provide drivers with higher driving comfort, smart vehicle interiors (such as seat orientation, etc.) have significantly changed compared with traditional vehicles. Therefore, the injury epidemiology and injury mechanism of drivers in intelligent vehicle accidents have new characteristics.

5. Passive safety technologies. The vehicle passive safety technologies should keep pace with the development of vehicle intelligence. In traditional vehicle accidents, passive safety research focuses on vehicle crashworthiness and airbag (curtain), etc. With the development of intelligent vehicle technologies, passive safety technologies should also become more intelligent. Analysing the effectiveness of new passive safety technologies in intelligent vehicle accidents and guiding their improvement have become important directions for future intelligent vehicle accident analysis.

5. Key issues of intelligent vehicle accident liability identification

The traditional accident liability identification adopts the ‘fault principle’, and there are corresponding solutions for those collisions between vehicles and other VRUs. However, intelligent vehicle accidents involve multiple parties and the liability boundaries between these parties are not clear, which may result in insufficient relief of the victim's rights. How to conduct liability identification for intelligent vehicle accidents has become an important issue that restricts the popularization and application of intelligent vehicles.

5.1 Status of foreign legislation on intelligent vehicles

As leaders in the development and production of traditional vehicles, western countries, such as the United States and Germany, have established relevant laws and regulations for intelligent vehicle technologies. However, there are differences between the United States and Germany regarding the identification of accident liability subject under autonomous driving.

American legislators believe that human drivers are the main initiators of vehicle accidents, so human drivers should be responsible for accidents. The intelligent vehicle is controlled by the system when it is in an autonomous driving state. Is the user still the driver? Do human drivers still have to shoulder the liability after an accident? Some states in the United States give a positive answer to this question. Legislators in California, Nevada and Florida believe that the person who actively turns on the automatic driving mode is the real driver of the intelligent vehicle, and they should be responsible for the consequences caused by the automatic driving [18].

The liability identification of intelligent vehicle accidents in Germany relies more on driving data. Some devices, such as ‘black boxes’, are necessary for the intelligent vehicles to record the detailed information at different stages, for example, the autonomous driving system is working, the human driver begins to intervene and the human driver is driving. In this way, the related liability belongs to the human driver if the accident occurs in a stage when he or she is driving. The intelligent vehicle manufacturer will take liability if the accident occurs in a stage when the autonomous driving system is working (causing an accident due to system failure, etc.) [19].

The states of foreign intelligent vehicle legislation are summarized in Table 3.

5.2 Liability identification for intelligent vehicle accidents in China

Detailed regulations for traditional road traffic accidents have been established in China, such as ‘Implementation Regulations of the Road Traffic Safety Law of the People's Republic of China’, ‘Procedures on Road Traffic Accident Handling Procedures’ and ‘Road Traffic Accident Handling Work Norms’. However, as far as intelligent vehicle accidents are concerned, there are still many problems, such as (1) a complete intelligent vehicle damage compensation and liability identification system has not been established in China; (2) related laws and regulations are not sound; and (3) many issues, such as the division of liability of subjects and the determination of compensation costs, are still unclear. This is mainly because there are two types of civil subjects in China, namely natural persons and legal persons. It also stipulates that only natural persons or legal persons with civil capacity can bear civil liability. The intelligent vehicle does not belong to either of the above two subjects, so it cannot serve as the liability subject.

The steps of determining intelligent vehicle accident liability contain, (1) clarifying the liability subjects; (2) classifying the liability subjects; and (3) confirming the applicable law. There are obvious differences about the liability subjects between intelligent vehicle accidents and traditional vehicle accidents. The liability subjects of intelligent vehicle accidents include vehicles, vehicle assistant drivers, manufacturers, vehicle owners and insurance companies. The liability nature of these five subjects can be divided into three categories, namely driving liability, management liability and insurance liability, as shown in Table 4. Specific liabilities and corresponding subjects can be determined in accordance with relevant laws. The division of intelligent vehicle accident liability should follow four principles, (1) the principle of behaviour liability; (2) the principle of causality; (3) the principle of right of way; and (4) the principle of safety [20, 21].

6. Analysis of a real intelligent vehicle accident

An intelligent vehicle accident reported on the Internet is shown in detail to explain how to conduct the accident cause analysis and the accident liability determination.

6.1 Case introduction

One day in 2020, an automated car collided with a truck that had rolled over previously due to the failure of the car's autopilot system. The car did not take any deceleration or steering action before the collision. According to the driver's recall, the driver activated the autopilot and set the driving speed at 110 km/h. The car driver thought the autopilot system could have detected obstacles and taken corresponding measures, but the actual situation was not the case.

6.2 Accident causation

In this accident, the driver did not take over the vehicle before the collision. One of the main causes is not using the driving assistance function correctly in accordance with the instructions. As far as intelligent vehicles are concerned, they use a combination of camera and millimetre wave radar to obtain external information, which is almost incompetent for nonstandard static objects. This because the camera needs to be developed by machine learning to recognize objects. However, static objects are so diverse and varied in shape, leading to the failure of the camera to recognize those objects that are not included in the existing samples. Although radar can identify static objects, the vehicle adopts a strategy of prioritizing visual perception. The camera often takes precedence over the radar in judgement. If the perception system makes a misjudgement, other systems may not respond. Therefore, the design flaw of the vehicle perception system is another major cause of the accident.

6.3 Accident liability identification

The user of the intelligent vehicle is the passenger, but the direct controller of the intelligent vehicle is the autonomous driving system made by the manufacturer. Based on the ‘Traffic Safety Law of the People's Republic of China’, accident liability can be determined as follows, for property loss caused by the accident, the driver should take the primary liability for not using the assistant driving function correctly in accordance with the instructions, and the intelligent vehicle should take secondary liability due to its technical defects.

7. Conclusions

Intelligence, automation and light weight are the inevitable directions of future automobile research and development. During the development of vehicles from nonautomated to fully automated, the road traffic system will be mixed with nonautomated vehicles and vehicles with different automation levels, making road traffic accidents more complex and diverse. Therefore, while indulging in the huge changes that intelligent vehicles have brought to social economy and human travel patterns, we also need to understand the safety issues caused by technical limitations and the social problems caused by imperfect laws and regulations.

Understanding the intelligent vehicle accident characteristics based on the existing intelligent vehicle accident data, proposing the key content of intelligent vehicle accident analysis based on the requirements of technology upgrades and safety enhancements, and establishing applicable intelligent vehicle accident investigation and identification methods can help researchers to understand the deficiencies of current intelligent vehicles more deeply, make improvements from the perspectives of perception, decision-making and execution, and build more scientific and complete ‘4E’ traffic safety intervention and improvement measures. Considering the complexity of future traffic accidents, it is suggested to establish relevant standards and regulations for intelligent vehicles as soon as possible, clarifying the legal liability borne by each liability subject in intelligent vehicle accidents, and putting forward management measures according to the actual requirements of accident prevention and handling.

All in all, despite the research and development of intelligent vehicles are obtaining more attention, there are still a large number of problems that need to be solved. It still takes time for smart transportation to be tested.

ACKNOWLEDGEMENTS

This study was supported by The National Natural Science Foundation of China under Grant No. 52072214.

Conflict of interest statement

None declared.

References