A method for constructing a library of hazardous test scenarios for autonomous driving based on driving behavior deviation

Abstract

This invention belongs to the field of autonomous vehicle testing and evaluation technology, specifically a method for constructing a hazardous test scenario library for autonomous vehicles based on driving behavior deviation. The method includes: setting a scenario extraction threshold; directly extracting hazardous test scenarios; analyzing the implicit risks of remaining safe scenarios; and eliminating hazardous scenarios from traffic vehicle driving safety behavior. This invention analyzes key scenarios neglected by traditional methods and transforms them into hazardous test scenarios through driving behavior deviation, fully utilizing natural driving data to construct a comprehensive hazardous test scenario library for autonomous vehicles.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle testing and evaluation technology, specifically a method for constructing a library of autonomous driving hazard test scenarios based on driving behavior deviations. Background Technology

[0002] Natural driving data is a crucial source for analyzing and extracting test scenarios for autonomous vehicles. However, most current methods only consider extracting key scenarios directly from natural driving data, neglecting the important information implicit in some seemingly unimportant test scenarios. Existing technologies extract dangerous test scenarios by setting behavioral indicator thresholds, ignoring natural driving data information below these thresholds, resulting in a significant waste of driving data. To address this, this invention proposes a method for generating dangerous test scenarios for autonomous vehicles based on driving behavior shifts. This method analyzes key scenarios neglected by traditional methods and transforms them into dangerous test scenarios through driving behavior shifts, fully utilizing natural driving data information to construct a comprehensive library of dangerous test scenarios for autonomous vehicles. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a method for constructing a hazardous test scenario library for autonomous driving based on driving behavior deviation. This method analyzes key scenarios neglected by traditional methods and transforms them into hazardous test scenarios through driving behavior deviation, making full use of natural driving data to construct a comprehensive hazardous test scenario library for autonomous vehicles.

[0004] The technical solution of this invention is described below in conjunction with the accompanying drawings:

[0005] A method for constructing a library of hazardous test scenarios for autonomous driving based on driving behavior deviation includes:

[0006] Set the scene extraction threshold;

[0007] Directly extract hazardous test scenarios;

[0008] Analyze the hidden risks in the remaining security scenarios;

[0009] Dangerous scenarios are eliminated from the safe driving behavior of vehicles.

[0010] Furthermore, the specific method for setting the scene extraction threshold is as follows:

[0011] The time to collision (TTC) and time to intrusion (PET) are used as direct extraction thresholds for hazardous test scenarios. After acquiring the movement trajectory of the vehicle and surrounding objects during natural driving, the TTC and PET between the vehicle and all surrounding objects are calculated in real time. TTC less than 2 seconds or PET less than 2 seconds are used as the extraction thresholds for hazardous scenarios.

[0012] TTC is used for road conditions without intersections; PET is used for intersection conditions; the time point at which both TTC and PET are met indicates the presence of hazardous scene data in the surrounding area; the calculation methods for TTC and PET are as follows:

[0013] TTC = |v ego -v target |×cosθ÷Δs

[0014] PET=|t target -t ego |

[0015] In the formula, v ego v is the velocity vector of this vehicle; target θ is the velocity vector of the surrounding target; θ is the angle between the direction of the velocity vector difference and the line connecting the center point of the vehicle and the target; Δs is the distance between the center point of the vehicle and the target; t target t is the time it takes for the target object to reach the target point at the intersection. ego This is the time it takes for this vehicle to reach the target point at the intersection.

[0016] Furthermore, the specific method for directly extracting hazardous test scenarios is as follows:

[0017] Search for real-time TTC and PET information corresponding to natural driving time series data. When TTC is less than 2 seconds or PET is less than 2 seconds, record the current time point and analyze the current road structure status, vehicle motion behavior status, and target object motion behavior status.

[0018] After acquiring the road structure status, vehicle motion behavior status, and target object motion behavior status, the road structure, vehicle motion behavior status, target object motion behavior status, vehicle position in the road, target object position in the road, and the relative position of the target object to the vehicle are encoded. The encoding order is: road structure, vehicle position in the road, vehicle motion behavior status, target object 1 position in the road, target object 1 relative position to the vehicle, target object 1 motion behavior status, target object 2 position in the road, target object 2 relative position to the vehicle, target object 2 motion behavior status, ... The number of target objects present corresponds to the number of target objects whose relevant information is encoded. Scene data with the same encoding structure are clustered to obtain the corresponding scene data for all scene types, thus initially constructing a natural driving hazard test scene library.

[0019] Furthermore, the road structure state is obtained through a neural network; the motion behavior state of the vehicle and the target object is obtained through high-dimensional time series data clustering.

[0020] Furthermore, the specific methods for analyzing the hidden risks in the remaining security scenarios are as follows:

[0021] After processing, the remaining natural driving data represents safe scenarios where real-time TTC and PET do not meet the threshold conditions. Implicit risk analysis is performed on the remaining safe scenarios. Based on the target object motion behavior state analysis method, the motion behavior state of all remaining natural driving data is analyzed, and a time series diagram of the motion behavior state corresponding to the target object is established. All non-uniform motion behavior states of surrounding targets under natural driving conditions are found, including acceleration, deceleration, turning, lane changing, and U-turn. These behaviors of the target object are replaced with the preceding motion behavior in sequence. That is, the preceding motion behavior is extended in time according to the state 2 seconds before the state transition point. Real-time TTC and PET are calculated for 10 seconds after each replacement node.

[0022] Furthermore, the specific method for temporally extending the preceding motion behavior according to the state 2 seconds before the state transition point is as follows:

[0023] Preceding behaviors include constant speed behavior, acceleration behavior, deceleration behavior, turning behavior, lane changing behavior, and U-turn behavior;

[0024] For uniform motion, the direction of the motion expansion is along the center line of the road, and the speed is the average speed of the two seconds before the expansion point.

[0025] For acceleration behavior, the direction of the behavior expansion motion is along the center line of the road, and the magnitude of the acceleration is the average acceleration in the 2 seconds before the expansion point;

[0026] For deceleration behavior, the direction of the behavior extension is along the center line of the road, the magnitude of the deceleration is the average deceleration in the 2 seconds before the extension point, and the final behavior of the deceleration behavior extension is to stop, without generating negative velocity.

[0027] For turning behavior, the direction of the extended movement is the direction of the arc with the current road centerline and the target lane road centerline as tangents. The speed trend of the extended turning behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. At the same time, the endpoint of the turning behavior is the direction of the target lane road centerline. When the target lane road centerline direction is reached, the turning behavior becomes a straight-line behavior.

[0028] For lane-changing behavior, the direction of the extended movement is a fifth-order polynomial with the current road centerline and the centerline of the outermost lane as tangents. The speed trend of the extended turning behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. At the same time, the endpoint of the lane-changing behavior is the direction of the centerline of the outermost lane. After reaching the direction of the centerline of the outermost lane, the lane-changing behavior becomes a straight-line behavior.

[0029] For the U-turn behavior, the direction of the extended movement is an arc with the current road centerline and the opposite road centerline as tangents. The speed trend of the extended U-turn behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. The speed magnitude refers to the constant speed, acceleration or deceleration extension method described above. At the same time, the endpoint of the U-turn behavior is the direction of the opposite lane road centerline. After reaching the direction of the opposite lane road centerline, the U-turn behavior becomes a straight-line behavior.

[0030] Furthermore, the specific methods for eliminating dangerous scenarios from traffic vehicle driving safety behavior are as follows:

[0031] Determine if a time-to-market (TTC) or time-to-petition (PET) value is less than 2 seconds and record the current time point. If so, it indicates that the behavior of the target object before replacement was a safe driving behavior, and removing it will result in a dangerous scenario. Encode the time points that meet the TTC and PET thresholds, along with the corresponding road structure, vehicle motion behavior state, target object motion behavior state, vehicle position on the road, target object position on the road, and the relative position of the target object and vehicle. For a given time point, there may be multiple surrounding target object behavior replacement operations. Each behavior replacement corresponds to TTC and PET calculations. If the preceding behavior extensions of different targets all lead to the generation of TTC or PET threshold conditions, then each preceding behavior extension corresponds to a dangerous test scenario. Compare the encoded dangerous test scenarios with the data in the dangerous test scenario library, and merge dangerous test scenarios with the same encoding to enrich the dangerous test scenario data extracted from natural driving data.

[0032] The beneficial effects of this invention are as follows:

[0033] This invention analyzes key scenarios that are overlooked by traditional methods and transforms them into dangerous test scenarios through driving behavior deviation, making full use of natural driving data information to build a comprehensive library of dangerous test scenarios for autonomous vehicles. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a schematic diagram of a method for constructing a hazardous test scenario library for autonomous driving based on driving behavior deviation, as described in this invention.

[0036] Figure 2 This is an example of the original natural driving process and a corresponding motion sequence diagram;

[0037] Figure 3 Example of the process of eliminating safe behaviors of vehicles and corresponding motion sequence diagram;

[0038] Figure 4 This is a TTC diagram illustrating the original vehicle driving process during natural driving.

[0039] Figure 5 This is a diagram illustrating the driving TTC after eliminating safe driving behaviors. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1

[0042] See Figure 1 This embodiment provides a method for constructing a library of autonomous driving hazard test scenarios based on driving behavior deviation, including:

[0043] S1. Set the scene extraction threshold, as follows:

[0044] The time to collision (TTC) and time to intrusion (PET) are used as direct extraction thresholds for hazardous test scenarios. After acquiring the movement trajectory of the vehicle and surrounding objects during natural driving, the TTC and PET between the vehicle and all surrounding objects (including vehicles and pedestrians) are calculated in real time. TTC less than 2 seconds (only considering cases greater than 0) or PET less than 2 seconds are used as hazardous scenario extraction thresholds.

[0045] TTC is used for road conditions without intersections; PET is used for intersection conditions (including intersections, ramp entrances / exits, and roundabout entrances / exits). The occurrence of a time point that satisfies both TTC and PET indicates the presence of hazardous scene data in the surrounding area. TTC and PET are calculated as follows:

[0046] TTC = |v ego -v target |×cosθ÷Δs

[0047] PET=|t target -t ego |

[0048] In the formula, v ego v is the velocity vector of this vehicle; targetθ is the velocity vector of the surrounding target; θ is the angle between the direction of the velocity vector difference and the line connecting the center point of the vehicle and the target; Δs is the distance between the center point of the vehicle and the target; t target t is the time it takes for the target object to reach the target point at the intersection. ego This is the time it takes for this vehicle to reach the target point at the intersection.

[0049] S2. Directly extract hazardous test scenarios, as follows:

[0050] Search for real-time TTC and PET information corresponding to natural driving time series data. When TTC is less than 2 seconds or PET is less than 2 seconds, record the current time point and analyze the current road structure status, vehicle motion behavior status, and target object motion behavior status.

[0051] The road structure status is obtained through neural networks, while the motion behavior status of the vehicle and the target object is obtained through high-dimensional time series data clustering.

[0052] After acquiring the road structure status, vehicle motion behavior status, and target object motion behavior status, the road structure, vehicle motion behavior status, target object motion behavior status, vehicle position in the road, target object position in the road, and the relative position of the target object to the vehicle are encoded. The encoding order is: road structure, vehicle position in the road, vehicle motion behavior status, target object 1 position in the road, target object 1 relative position to the vehicle, target object 1 motion behavior status, target object 2 position in the road, target object 2 relative position to the vehicle, target object 2 motion behavior status, ... The number of target objects present corresponds to the number of target objects whose relevant information is encoded. Scene data with the same encoding structure are clustered to obtain the corresponding scene data for all scene types, thus initially constructing a natural driving hazard test scene library.

[0053] S3. Analyze the hidden risks in the remaining security scenarios, as follows:

[0054] See Figure 2 After processing by S1 and S2, the remaining natural driving data represents safe scenarios where real-time TTC and PET do not meet the threshold conditions. Implicit risk analysis is then performed on these remaining safe scenarios. Based on the target object motion behavior state analysis method from the second step, the motion behavior states of all remaining natural driving data are analyzed, and a time series diagram of the target object's motion behavior state is established. All non-uniform motion behavior states of surrounding targets under natural driving conditions are identified, including acceleration, deceleration, turning, lane changing, and U-turns. These behaviors of the target object are then replaced with their preceding motion behaviors, i.e., the preceding motion behaviors are time-series extended according to the state 2 seconds before the state transition point. Real-time TTC and PET are calculated for 10 seconds after each replacement node.

[0055] The specific methods for extending the preceding behavior are as follows:

[0056] Preceding behaviors include constant speed behavior, acceleration behavior, deceleration behavior, turning behavior, lane changing behavior, and U-turn behavior.

[0057] For uniform motion, the direction of the motion is along the center line of the road, and the speed is the average speed of the two seconds before the point of expansion; for accelerated motion, the direction of the motion is along the center line of the road, and the acceleration is the average acceleration of the two seconds before the point of expansion.

[0058] For deceleration behavior, the direction of its extended motion is along the center line of the road, and the magnitude of the deceleration is the average deceleration in the two seconds before the extension point. It should be noted that the final behavior of the deceleration behavior extension is to come to a standstill and does not produce negative velocity.

[0059] For turning behavior, the direction of its extended movement is an arc with the current road centerline and the target lane road centerline as tangents. It should be noted that the speed trend of the extended turning behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. The speed magnitude refers to the constant speed, acceleration or deceleration extension method described above. At the same time, the endpoint of the turning behavior is the direction of the target lane road centerline. When the target lane road centerline direction is reached, the turning behavior becomes a straight-line behavior.

[0060] For lane-changing behavior, the direction of its extended movement is a fifth-order polynomial with the current road centerline and the centerline of the outermost lane as tangents. It should be noted that the speed trend of the extended turning behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. The speed magnitude refers to the constant speed, acceleration or deceleration extension method described above. At the same time, the endpoint of the lane-changing behavior is the direction of the centerline of the outermost lane. When the direction of the centerline of the outermost lane is reached, the lane-changing behavior becomes a straight-line behavior.

[0061] For the U-turn behavior, the direction of its extended movement is an arc with the current road centerline and the opposite road centerline as tangents. It should be noted that the speed trend of the U-turn behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. The speed magnitude refers to the constant speed, acceleration or deceleration extension method described above. At the same time, the endpoint of the U-turn behavior is the direction of the opposite lane road centerline. After reaching the direction of the opposite lane road centerline, the U-turn behavior becomes a straight line behavior.

[0062] S4. Eliminate dangerous scenarios from the generation of safe driving behaviors in traffic vehicles, as detailed below:

[0063] See Figure 3The process begins by determining whether a Time To Completion (TTC) or Petition Detection (PET) value of less than 2 seconds occurs in step S3, and recording the current time point. If this occurs, it indicates that the target object's behavior before replacement was a safe driving behavior, and removing it would create a dangerous scenario. The time points that satisfy the TTC and PET thresholds, along with the corresponding road structure, vehicle motion behavior state, target object motion behavior state, vehicle's position on the road, target object's position on the road, and the relative position of the target object and vehicle, are encoded. This step follows the method described in step S2. It's important to note that the target vehicle's behavior is an extension of the preceding motion behavior in step three, not the original motion behavior from the natural driving data. For a given time point, there may be multiple surrounding target object behavior replacement operations. Each behavior replacement corresponds to TTC and PET calculations. If the preceding behavior extensions of different targets all lead to the TTC or PET threshold conditions, then each preceding behavior extension corresponds to a dangerous test scenario. The encoded dangerous test scenarios are compared with the data in the dangerous test scenario library from step two. Dangerous test scenarios with the same encoding are merged, thereby enriching the dangerous test scenario data extracted from the natural driving data.

[0064] Ultimately, S2 and S4 jointly generate a hazardous scenario library. This invention analyzes and extracts some scenarios that cannot be extracted using traditional methods, thus constructing a comprehensive hazardous test scenario library for autonomous vehicles.

[0065] Example 2

[0066] This embodiment experimentally verifies the method for constructing an autonomous driving hazard test scenario library based on driving behavior deviation provided in Embodiment 1, as detailed below:

[0067] The collected natural driving data is used to extract test scenarios. The natural driving data collection vehicle uses a horizon camera to analyze the driving trajectories of surrounding traffic vehicles, and finally obtains an autonomous driving database covering the movement trajectories of the vehicle and other vehicles.

[0068] Step S2 directly extracts scene thresholds during driving, resulting in 15 hazardous test scenarios, including 3 scenarios involving the vehicle ahead, 6 scenarios involving the vehicle ahead braking, and 6 scenarios involving intersections. Steps S3 and S4 of this invention are then used to extract the remaining scenarios, yielding 24 scenarios from the remaining natural driving data, greatly enriching the types of test scenario databases. Figure 4 and Figure 5 This refers to the real-time TTC of a certain segment of driving data during the analysis process using the method of this invention. Figure 4 The original vehicle driving process during natural driving (TTC) is the process of driving the vehicle naturally. Figure 5The two figures show the driving TTC after removing vehicle safety behaviors. A comparison reveals that the method of this invention can effectively identify hidden dangerous scenarios during natural driving, extracting hidden scenario information from driving data that traditional methods cannot detect, thus significantly improving the richness of the extracted test scenarios.

[0069] In summary, this invention can analyze key scenarios that are overlooked by traditional methods, and transform them into dangerous test scenarios through driving behavior deviations. It makes full use of natural driving data information and ultimately builds a comprehensive library of dangerous test scenarios for autonomous vehicles.

[0070] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the scope of protection of the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, any person skilled in the art can make equivalent substitutions or changes based on the technical solution and inventive concept of the present invention within the scope of the technology disclosed in the present invention. These simple modifications are all within the scope of protection of the present invention.

[0071] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. In order to avoid unnecessary repetition, the present invention will not describe the various possible combinations separately.

[0072] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed by the present invention.

Claims

1. A method for constructing a library of hazardous test scenarios for autonomous driving based on driving behavior deviation, characterized in that, include: Set the scene extraction threshold, as follows: The collision time (TTC) and post-intrusion time (PET) are used as direct extraction thresholds for dangerous test scenarios. After obtaining the movement trajectory of the vehicle and surrounding objects during natural driving, the TTC and PET between the vehicle and all surrounding objects are calculated in real time. The threshold for dangerous scenarios is set as TTC less than 2 seconds or PET less than 2 seconds. Among them, TTC is used for road conditions without intersections; PET is used for intersection conditions; the time point that meets both TTC and PET is the data of a dangerous scene in the surrounding area; Directly extract hazardous test scenarios; The following methods are used to analyze the hidden risks in the remaining security scenarios: After processing, the remaining natural driving data are safety scenarios where real-time TTC and PET do not meet the threshold conditions. Implicit risk analysis is performed on the remaining safety scenarios. Based on the target object motion behavior state analysis method, the motion behavior state of the target objects in all remaining natural driving data is analyzed, and a time series diagram of the motion behavior state of the target objects is established. All non-uniform motion behavior states of the surrounding target objects under the natural driving state are found, including acceleration, deceleration, turning, lane changing, and U-turn. These behaviors of the target objects are replaced with the preceding motion behaviors in turn. That is, the preceding motion behaviors are extended in time according to the state 2 seconds before the state transition point. The real-time TTC and PET are calculated 10 seconds after the replacement node after each replacement. The specific method for temporally extending the preceding motion behavior according to the state 2 seconds before the state transition point is as follows: Preceding behaviors include constant speed behavior, acceleration behavior, deceleration behavior, turning behavior, lane changing behavior, and U-turn behavior; For uniform motion, the direction of the motion expansion is along the center line of the road, and the speed is the average speed of the two seconds before the expansion point. For acceleration behavior, the direction of the behavior expansion motion is along the center line of the road, and the magnitude of the acceleration is the average acceleration in the 2 seconds before the expansion point; For deceleration behavior, the direction of the behavior extension is along the center line of the road, the magnitude of the deceleration is the average deceleration in the 2 seconds before the extension point, and the final behavior of the deceleration behavior extension is to stop, without generating negative velocity. For turning behavior, the direction of the extended movement is the direction of the arc with the current road centerline and the target lane road centerline as tangents. The speed trend of the extended turning behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. At the same time, the endpoint of the turning behavior is the direction of the target lane road centerline. When the target lane road centerline direction is reached, the turning behavior becomes a straight-line behavior. For lane-changing behavior, the direction of the behavior extension is a fifth-order polynomial with the current road centerline and the centerline of the outermost lane as tangents. The speed trend of the lane-changing behavior extension is consistent with the previous behavior, including constant speed, acceleration or deceleration. At the same time, the endpoint of the lane-changing behavior is the direction of the centerline of the outermost lane. When the direction of the centerline of the outermost lane is reached, the lane-changing behavior becomes a straight-line behavior. For the U-turn behavior, the direction of the extended movement is an arc with the current road centerline and the opposite road centerline as tangents. The speed trend of the extended U-turn behavior is consistent with the previous behavior, including constant speed, acceleration or deceleration. The speed magnitude is the same as the constant speed, acceleration or deceleration extension method. At the same time, the endpoint of the U-turn behavior is the opposite lane road centerline direction. When the opposite lane road centerline direction is reached, the U-turn behavior becomes a straight line behavior. The specific method for eliminating dangerous scenarios from traffic vehicle driving safety behavior is as follows: Determine if a Time Tolerance (TTC) or Petition Parameter (PET) value is less than 2 seconds and record the current time point. If so, it indicates that the behavior of the target object before replacement was a safe driving behavior, and removing it will result in a dangerous scenario. Encode the time points that meet the TTC and PET thresholds, along with the corresponding road structure, vehicle motion behavior state, target object motion behavior state, vehicle position on the road, target object position on the road, and the relative position of the target object and vehicle. For a given moment, there may be multiple surrounding target object behavior replacement operations. Each behavior replacement corresponds to TTC and PET calculations. If the preceding behavior extensions of different targets all lead to the generation of TTC or PET threshold conditions, then each preceding behavior extension corresponds to a dangerous test scenario. Compare the encoded dangerous test scenarios with the data in the dangerous test scenario library, and merge dangerous test scenarios with the same encoding to enrich the dangerous test scenario data extracted from natural driving data.

2. The method for constructing an autonomous driving hazard test scenario library based on driving behavior deviation according to claim 1, characterized in that, The calculation methods for TTC and PET are as follows: In the formula, This is the velocity vector of the vehicle; The velocity vector of the surrounding target objects; The angle between the direction of the velocity vector difference and the line connecting the center point of the vehicle and the target object; The distance between the center point of the transport vehicle and the center point of the target object; The time it takes for the target object to reach the target point at the intersection; This is the time it takes for this vehicle to reach the target point at the intersection.

3. The method for constructing an autonomous driving hazard test scenario library based on driving behavior deviation according to claim 1, characterized in that, The specific method for directly extracting hazardous test scenarios is as follows: Search for real-time TTC and PET information corresponding to natural driving time series data. When TTC is less than 2 seconds or PET is less than 2 seconds, record the current time point and analyze the current road structure status, vehicle motion behavior status, and target object motion behavior status. After acquiring the road structure status, vehicle motion behavior status, and target object motion behavior status, the road structure, vehicle motion behavior status, target object motion behavior status, vehicle position in the road, target object position in the road, and the relative position of the target object to the vehicle are encoded. The encoding order is: road structure, vehicle position in the road, vehicle motion behavior status, target object 1 position in the road, target object 1 relative position to the vehicle, target object 1 motion behavior status, target object 2 position in the road, target object 2 relative position to the vehicle, target object 2 motion behavior status, ... The number of target objects present corresponds to the number of target objects whose relevant information is encoded. Scene data with the same encoding structure are clustered to obtain the corresponding scene data for all scene types, thus initially constructing a natural driving hazard test scene library.

4. The method for constructing an autonomous driving hazard test scenario library based on driving behavior deviation according to claim 3, characterized in that, The road structure state is obtained through a neural network; the motion behavior state of the vehicle and the target object is obtained through high-dimensional time series data clustering.

Images