Automatic driving multi-vehicle dangerous interaction test scene generation method based on diffusion model

Abstract

The present application belongs to the technical field of automatic driving car testing, and in particular to a kind of automatic driving multi-vehicle dangerous interaction test scene generation method based on diffusion model. First, the multi-vehicle interaction process is represented as discrete sequence data, and is divided into known collision terminal state and to-be-generated trajectory segment;Subsequently, under the constraint that the initial state of driving is stable, the data before collision terminal state is completed in reverse through deterministic diffusion model;Guidance mechanism is added in the reverse completion process, and the stability, safety, rationality and collision consistency of trajectory generation are constrained;After the sampling trajectory is completed, the final trajectory generated by the bicycle model is physically corrected;Based on the obtained multi-vehicle interaction trajectory, the test scene of automatic driving car is converted through scene element completion. The present application can improve the efficiency of dangerous scene generation, reduce the waste of computing power, and ultimately establish a complete dangerous scene database to accelerate the testing and verification of automatic driving car.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle testing technology, specifically a method for generating autonomous driving multi-vehicle dangerous interaction test scenarios based on a diffusion model. Background Technology

[0002] Scenario-based testing methods have become the mainstream for testing and verifying autonomous vehicles, with the core being the construction of a test scenario library covering a wide range of scenario types. my country's "Notice on Carrying Out Pilot Work on Access and Road Traffic for Intelligent Connected Vehicles" explicitly states that autonomous vehicle test scenario types should include nominal scenarios, hazardous scenarios, and edge scenarios. Therefore, how to generate hazardous test scenarios in batches has become a core task in the generation of autonomous vehicle test scenarios. Currently, the hazardous scenario generation process often adopts steady-state, interactive, and hazardous behavior flows based on vehicle motion states. Methods such as reinforcement learning and generative networks are used to start from the vehicle's steady-state state, design vehicle motion interactions, and ultimately guide the occurrence of the final hazard. Due to the high-dimensional characteristics of vehicle temporal motion behavior, this forward sampling method often results in safe scenarios lacking hazardous interactions. Although a certain number of hazardous scenarios can be obtained through screening, it involves a significant waste of computational resources. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a method for generating hazardous interaction test scenarios for autonomous driving based on a diffusion model. This method introduces a diffusion model into the multi-vehicle interaction trajectory generation process, using vehicle dynamics as constraints and a guided deterministic diffusion model to progressively generate hazardous interaction scenarios that conform to natural driving principles. First, the multi-vehicle interaction process is represented as discrete sequence data and divided into known collision final states and trajectory segments to be generated. Then, under the constraint of a stable initial driving state, the data before the collision final state is reverse-completed using a deterministic diffusion model. A guiding mechanism is added during the reverse-complete process to constrain the stability, safety, rationality, and collision consistency of trajectory generation. After the sampled trajectory is completed, a bicycle model is used to physically correct the final sampled trajectory. Based on the obtained multi-vehicle interaction trajectory, it is transformed into a test scenario for autonomous vehicles through scene element completion. The results of this invention can help enterprises establish an efficient collision hazard scenario generation system, improve the efficiency of hazard scenario generation, reduce computational waste, and ultimately establish a complete hazard scenario database, accelerating the testing and verification of autonomous vehicles.

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

[0005] This invention provides a method for generating hazardous interaction test scenarios for autonomous driving multi-vehicles based on a diffusion model, comprising the following steps:

[0006] Step 1: Represent the multi-vehicle interaction process as discrete sequence data and divide it into known collision final states and trajectory segments to be generated;

[0007] Step 2: Under the constraint that the initial driving state is stable, the data before the final collision state is back-completed using a deterministic diffusion model;

[0008] Step 3: Incorporate a guidance mechanism into the reverse completion process to constrain the stability, security, rationality, and collision consistency of trajectory generation;

[0009] Step 4: After the sampling trajectory is completed, use a bicycle model to physically correct the final trajectory generated by the sampling.

[0010] Step 5: Based on the obtained multi-vehicle interaction trajectories, the scenarios are transformed into test scenarios for autonomous vehicles by completing the scene elements.

[0011] Furthermore, the specific method for step one is as follows:

[0012] 11. Assume the scene contains If there are 10 vehicles, then the set of vehicles is represented as:

[0013] ;

[0014] Vehicle information is described using the Frenet coordinate system, for any vehicle... In time step The driving status at this location is represented as follows:

[0015] ;

[0016] In the formula, This refers to the longitudinal position of the vehicle along the centerline of the road. The lateral offset of the vehicle relative to the center line of the lane; The deviation angle between the vehicle's heading and the road direction; The vehicle's speed;

[0017] At time step At this point, the joint motion state of multiple vehicles is represented as:

[0018] ;

[0019] Over the complete time step, the multi-vehicle interaction trajectory segment is represented as follows:

[0020] ;

[0021] 12. Divide the trajectory sequence into a collision final state and a trajectory segment to be generated. The collision final state is the unmodifiable part of the trajectory, while the trajectory segment to be generated is the modifiable part. Use a time masking function to distinguish between them. Let the length of the trajectory segment in the final collision state be... The masking function is then defined as:

[0022] ;

[0023] In the formula, For the total time series length of the trajectory, when =1 indicates that the trajectory of the current part can be modified; at the same time, the tensor of the final state trajectory segment of the terminal collision is defined in order to carry out the subsequent diffusion model generation process.

[0024] ;

[0025] 13. During the interactive trajectory generation process, it is stipulated that the initial state of the trajectory segment to be generated must meet the stable driving conditions, as shown in the following formula:

[0026] ;

[0027] The set of stable driving states Defined as:

[0028] .

[0029] Furthermore, the specific method for step two is as follows:

[0030] 21. Perform forward diffusion on the deterministic diffusion model; define the diffusion noise scheduling parameter as follows: Meanwhile, the parameter changes during the diffusion process are denoted as:

[0031] ;

[0032] In the formula, This is the noise accumulation term; For the first The noise scheduling parameters for each diffusion step are used to control the relative proportion of the original trajectory information and the noise component in that step.

[0033] At this point, the forward diffusion process of the entire trajectory generation process diffusion model can be represented as:

[0034] ;

[0035] In the formula, For the first Noisy trajectory; It is Gaussian noise;

[0036] 22. After the forward diffusion process of the diffusion model is defined in step 21, the backsampling process is designed; in the generation phase, Gaussian noise is used to initialize the entire trajectory.

[0037] ;

[0038] In the formula, It is a Gaussian process;

[0039] In the backsampling process, that is, from the first... Step towards During the -1 step sampling process, the predicted noise term of the diffusion model is:

[0040] ;

[0041] In the formula, For scene condition information related to trajectory generation; For parameters The noise prediction function output by the diffusion model; Masking time;

[0042] At this point, the calculated noiseless trajectory estimate is expressed by the following formula:

[0043] ;

[0044] Based on the update rule of the deterministic diffusion model, we obtain Trajectory information at time step -1:

[0045] ;

[0046] In the formula, To represent the update rule based on the deterministic diffusion implicit model at the 1st... Trajectory estimation results obtained by -1 backsampling time steps; In the diffusion model, the first -1 is the cumulative product of the noise scheduling parameters corresponding to the sampling steps.

[0047] Furthermore, the specific method for step three is as follows:

[0048] 31. Four indicators—stability, safety, rationality, and collision consistency—are introduced to constrain the trajectory generation process. The constructed trajectory guidance cost function is as follows:

[0049] ;

[0050] In the formula, , , , These are the costs of stability, safety, rationality, and collision consistency constraints, respectively.

[0051] 32. After step 22 is completed, the cost function established in step 31 is introduced to guide and correct the trajectory generated in step 22:

[0052] ;

[0053] In the formula, Guided weights related to the number of sampling steps; The gradient of the cost function with respect to the trajectory;

[0054] 33. After completing the guiding correction, perform the end projection operation of the diffusion model to ensure that all collision final states satisfy the collision constraints:

[0055] ;

[0056] In the formula, Represents the element-wise operator.

[0057] Furthermore, the specific method for step four is as follows:

[0058] 41. For each vehicle, a separate bicycle model is introduced for calibration. The state update relationship is as follows:

[0059] ;

[0060] ;

[0061] ;

[0062] ;

[0063] In the formula, This refers to the vehicle's wheelbase. For the first The car in time The longitudinal position along the centerline of the road; The discrete time interval between adjacent time steps; For the first The car in time Lateral offset along the road centerline; For the first The car in time Front wheel angle at the time; For the first The car in time The longitudinal acceleration at that time.

[0064] Furthermore, the specific method for step five is as follows:

[0065] 51. Based on the generated vehicle interaction trajectory, and using a six-layer scenario model, the four scenario elements of weather, road facilities, vehicle type, and road condition are completed through a combination test method. At the same time, the complete scenario description is described using the OpenScenario format to achieve standardized generation and storage of scenarios.

[0066] The beneficial effects of this invention are as follows:

[0067] The results of this invention can help enterprises establish an efficient collision hazard scenario generation system, improve the efficiency of hazard scenario generation, reduce computing power waste, and ultimately establish a complete hazard scenario database to accelerate the testing and verification of autonomous vehicles. Attached Figure Description

[0068] 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.

[0069] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0071] See Figure 1 This invention provides a method for generating dangerous interaction test scenarios for autonomous driving multi-vehicles based on a diffusion model, comprising the following steps:

[0072] Step 1: Represent the multi-vehicle interaction process as discrete sequence data, and divide it into known collision final states and trajectory segments to be generated. The specific method is as follows:

[0073] 11. Assume the scene contains If there are 10 vehicles, then the set of vehicles is represented as:

[0074] ;

[0075] Vehicle information is described using the Frenet coordinate system, for any vehicle... In time step The driving status at this location can be represented as:

[0076] ;

[0077] In the formula, This refers to the longitudinal position of the vehicle along the centerline of the road. This refers to the lateral offset of the vehicle relative to the center line of the lane. This is the deviation angle between the vehicle's heading and the road direction. This refers to the vehicle's speed.

[0078] At time step At this point, the combined motion of multiple vehicles can be represented as:

[0079] ;

[0080] Over the complete time step, the multi-vehicle interaction trajectory segment can be represented as:

[0081] ;

[0082] 12. Divide the trajectory sequence into a collision final state and a trajectory segment to be generated. The collision final state is the unmodifiable part of the trajectory, while the trajectory segment to be generated is the modifiable part. Use a time masking function to distinguish between them. Let the length of the trajectory segment in the final collision state be... The masking function is then defined as:

[0083] ;

[0084] In the formula, For the total time series length of the trajectory, when =1 indicates that the trajectory of the current part can be modified; at the same time, the tensor of the final state trajectory segment of the terminal collision is defined in order to carry out the subsequent diffusion model generation process.

[0085] ;

[0086] 13. During the interactive trajectory generation process, it is stipulated that the initial state of the trajectory segment to be generated must meet the stable driving conditions, as shown in the following formula:

[0087] ;

[0088] The set of stable driving states Defined as:

[0089] .

[0090] Step 2: Under the constraint that the initial driving state is stable, the data before the final collision state is back-completed using a deterministic diffusion model. The specific method is as follows:

[0091] 21. Perform forward diffusion on the deterministic diffusion model; define the diffusion noise scheduling parameter as follows: Meanwhile, the parameter changes during the diffusion process are denoted as:

[0092] ;

[0093] In the formula, This is the noise accumulation term; For the first The noise scheduling parameters for each diffusion step are used to control the relative proportion of the original trajectory information and the noise component in that step.

[0094] At this point, the forward diffusion process of the entire trajectory generation process diffusion model can be represented as:

[0095] ;

[0096] In the formula, For the first Noisy trajectory; It is Gaussian noise;

[0097] 22. After defining the forward diffusion process of the diffusion model (i.e., step 21), design its backsampling process; in the generation phase, use Gaussian noise to initialize the entire trajectory:

[0098] ;

[0099] In the formula, It is a Gaussian process;

[0100] In the backsampling process, that is, from the first... Step towards During the -1 step sampling process, the predicted noise term of the diffusion model is:

[0101] ;

[0102] In the formula, For scene condition information related to trajectory generation; For parameters The noise prediction function output by the diffusion model; Masking time;

[0103] At this point, the calculated noiseless trajectory estimate is expressed by the following formula:

[0104] ;

[0105] Based on the update rule of the deterministic diffusion model, we obtain Trajectory information at time step -1:

[0106] ;

[0107] In the formula, To represent the update rule based on the deterministic diffusion implicit model at the 1st... Trajectory estimation results obtained by -1 backsampling time steps; In the diffusion model, the first -1 is the cumulative product of the noise scheduling parameters corresponding to the sampling steps.

[0108] Step 3: Incorporate a guidance mechanism into the reverse completion process to constrain the stability, safety, rationality, and collision consistency of trajectory generation. The specific method is as follows:

[0109] 31. Due to the inherent randomness of the diffusion model, four indicators—stability, safety, rationality, and collision consistency—are introduced to constrain the trajectory generation process in order to ensure the rationality of trajectory generation. The constructed trajectory guidance cost function is as follows:

[0110] ;

[0111] In the formula, , , , These are the costs of stability, safety, rationality, and collision consistency constraints, respectively.

[0112] 32. After step 22 is completed, the cost function established in step 31 is introduced to guide and correct the trajectory generated in step 22:

[0113] ;

[0114] In the formula, Guided weights related to the number of sampling steps; The gradient of the cost function with respect to the trajectory;

[0115] 33. After completing the guiding correction, perform the end projection operation of the diffusion model to ensure that all collision final states satisfy the collision constraints:

[0116] ;

[0117] In the formula, Represents the element-wise operator.

[0118] Step 4: After the sampling trajectory is completed, use a bicycle model to perform physical correction on the final trajectory generated by the sampling. The specific method is as follows:

[0119] 41. For each interactive trajectory generated by the guiding deterministic diffusion model, although its generation process satisfies the mandatory constraints of final-state collision and initial-state stability, the use of a cost function to constrain the intermediate trajectory generation process may produce some behavioral states that do not conform to vehicle dynamics. Therefore, it is necessary to perform certain physical corrections on the generated final trajectory. In this invention, a bicycle model is introduced for correction for each vehicle, and its state update relationship is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] ;

[0124] In the formula, This refers to the vehicle's wheelbase. For the first The car in time The longitudinal position along the centerline of the road; The discrete time interval between adjacent time steps; For the first The car in time Lateral offset along the road centerline; For the first The car in time Front wheel angle at the time; For the first The car in time The longitudinal acceleration at that time.

[0125] Step 5: Based on the obtained multi-vehicle interaction trajectories, convert them into test scenarios for autonomous vehicles by completing scene elements. The specific method is as follows:

[0126] 51. Based on the generated vehicle interaction trajectory, and using a six-layer scenario model, the four scenario elements of weather, road facilities, vehicle type, and road condition are completed through a combination test method. At the same time, the complete scenario description is described using the OpenScenario format to achieve standardized generation and storage of scenarios.

[0127] In summary, the results of this invention can help enterprises establish an efficient collision hazard scenario generation system, improve the efficiency of hazard scenario generation, reduce computing power waste, and ultimately establish a complete hazard scenario database to accelerate the testing and verification of autonomous vehicles.

[0128] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating hazardous interaction test scenarios for autonomous driving multi-vehicles based on a diffusion model, characterized in that, Includes the following steps: Step 1: Represent the multi-vehicle interaction process as discrete sequence data and divide it into known collision final states and trajectory segments to be generated; Step 2: Under the constraint that the initial driving state is stable, the data before the final collision state is back-completed using a deterministic diffusion model. The specific method is as follows: S21. Perform forward diffusion on the deterministic diffusion model; define the diffusion noise scheduling parameter as follows: Meanwhile, the parameter changes during the diffusion process are denoted as: ； In the formula, This is the noise accumulation term; For the first The noise scheduling parameters for each diffusion step are used to control the relative proportion of the original trajectory information and the noise component in that step. At this point, the forward diffusion process of the entire trajectory generation process diffusion model can be represented as: ； In the formula, For the first Noisy trajectory; It is Gaussian noise; S22. After the forward diffusion process of the diffusion model is defined (i.e., step S21), the backsampling process is designed; in the generation phase, Gaussian noise is used to initialize the entire trajectory. ； In the formula, It is a Gaussian process; In the backsampling process, that is, from the first... Step towards During the -1 step sampling process, the predicted noise term of the diffusion model is: ； In the formula, For scene condition information related to trajectory generation; For parameters The noise prediction function output by the diffusion model; Masking time; At this point, the calculated noiseless trajectory estimate is expressed by the following formula: ； According to the update rule of the deterministic diffusion model, we get Trajectory information at time step -1: ； In the formula, To represent the update rule based on the deterministic diffusion implicit model at the 1st... Trajectory estimation results obtained by -1 backsampling time steps; In the diffusion model, the first -1 cumulative product of noise scheduling parameters corresponding to sampling steps; Step 3: Incorporate a guidance mechanism into the reverse completion process to constrain the stability, safety, rationality, and collision consistency of trajectory generation. The specific method is as follows: S31. Four indicators—stability, safety, rationality, and collision consistency—are introduced to constrain the trajectory generation process. The constructed trajectory guidance cost function is as follows: ； In the formula, , , , These are the costs of stability, safety, rationality, and collision consistency constraints, respectively. S32. After step S22 is completed, the cost function established in step S31 is introduced to guide and correct the trajectory generated in step 22: ； In the formula, Guided weights related to the number of sampling steps; The gradient of the cost function with respect to the trajectory; S33. After completing the guiding correction, perform the end projection operation of the diffusion model to ensure that all collision final states satisfy the collision constraints: ； In the formula, Represents the element-wise operator; Step 4: After the sampling trajectory is completed, use a bicycle model to physically correct the final trajectory generated by the sampling. Step 5: Based on the obtained multi-vehicle interaction trajectories, transform them into test scenarios for autonomous vehicles by completing scene elements.

2. The method for generating autonomous driving multi-vehicle dangerous interaction test scenarios based on a diffusion model according to claim 1, characterized in that, The specific method for step one is as follows: S11, Assume the scene contains If there are 10 vehicles, then the set of vehicles is represented as: ； Vehicle information is described using the Frenet coordinate system, for any vehicle... In time step The driving status at this location is represented as follows: ； In the formula, This refers to the longitudinal position of the vehicle along the centerline of the road. The lateral offset of the vehicle relative to the center line of the lane; The deviation angle between the vehicle's heading and the road direction; The vehicle's speed; At time step At this point, the joint motion state of multiple vehicles is represented as: ； Over the complete time step, the multi-vehicle interaction trajectory segment is represented as follows: ； S12. Divide the trajectory sequence into the collision final state and the trajectory segment to be generated. The collision final state is the unmodifiable trajectory part, and the trajectory segment to be generated is the modifiable generation part. Use a time masking function to distinguish them. Let the length of the trajectory segment in the final collision state be... The masking function is then defined as: ； In the formula, For the total time series length of the trajectory, when =1 indicates that the trajectory of the current part can be modified; at the same time, the tensor of the final state trajectory segment of the terminal collision is defined in order to carry out the subsequent diffusion model generation process. ； S13. During the interactive trajectory generation process, it is stipulated that the initial state of the trajectory segment to be generated must meet the stable driving conditions, as shown in the following formula: ； The set of stable driving states Defined as: 。 3. The method for generating autonomous driving multi-vehicle hazardous interaction test scenarios based on a diffusion model according to claim 1, characterized in that, The specific method for step four is as follows: S41. For each vehicle, a separate bicycle model is introduced for calibration. The state update relationship is as follows: ； ； ； ； In the formula, This refers to the vehicle's wheelbase. For the first The car in time The longitudinal position along the centerline of the road; The discrete time interval between adjacent time steps; For the first The car in time Lateral offset along the road centerline; For the first The car in time Front wheel angle at the time; For the first The car in time The longitudinal acceleration at that time.

4. The method for generating autonomous driving multi-vehicle dangerous interaction test scenarios based on a diffusion model according to claim 1, characterized in that, The specific method for step five is as follows: S51. Based on the generated vehicle interaction trajectory, and based on the six-layer scenario model, the four scenario elements of weather, road facilities, vehicle type, and road condition are completed through a combination test method. At the same time, the complete scenario description is described using the OpenScenario format to achieve standardized generation and storage of scenarios.

Images