Method for generating humanized trajectory of background vehicle facing automatic driving test scene

Abstract

The application discloses a background vehicle personification trajectory generation method for an automatic driving test scene, and comprises the following steps: extracting driving features in a plurality of interactive scenes according to a set interactive logic to perform simulation training, thereby constructing a driving simulator; and taking the driving features in each interactive scene as driving behavior data of a plurality of background vehicles, configuring the driving behavior data with a set driving style one by one, setting a corresponding driving strategy for each driving style to construct a dynamic Bayesian game framework; obtaining the driving style of the background vehicles in any interactive scene in real time, performing iterative training on a multi-objective utility function, calculating the corresponding income value of the driving style of the background vehicles by using the multi-objective utility function, selecting a corresponding driving strategy from the dynamic Bayesian game framework to generate a trajectory control instruction, and performing personification optimization on the trajectory control instruction by using a random disturbance mechanism, so that multi-dimensional driving behavior dynamic generation and complex interactive logic accurate modeling are realized.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving testing technology, specifically to a method for generating anthropomorphic trajectories of background vehicles for autonomous driving testing scenarios. Background Technology

[0002] In recent years, with the continuous promotion of autonomous driving technology, the number of autonomous vehicles on the road has been increasing. This has led to a number of accidents involving autonomous vehicles, exposing the shortcomings of current autonomous driving systems in handling complex traffic scenarios and highlighting the need for strengthened testing. Current testing methods mainly include virtual simulation testing, real-vehicle road testing, and hybrid virtual-real testing. While real-vehicle road testing can directly evaluate autonomous vehicles in real traffic environments, it suffers from two major drawbacks: long testing cycles and high costs. To avoid the shortcomings of real-vehicle road testing, virtual simulation testing and hybrid virtual-real testing have gradually become the main methods for autonomous driving testing. Both of these testing methods are conducted in simulated testing scenarios, significantly reducing testing costs and time, improving testing efficiency, and enabling the tested vehicle to complete a large number of complex scenario tests in a short period of time, thus gaining widespread application.

[0003] Currently, in simulation testing scenarios, the vehicle under test needs to interact with background vehicles in the scenario to evaluate its driving performance in dynamic scenarios such as lane changing, following, and overtaking. Therefore, the behavior of the background vehicles has a significant impact on the performance evaluation of autonomous driving systems. For example, the closer the behavior of the background vehicles is to the intentions of a real driver, the higher the realism of the simulation testing scenario, and the more accurate the test results. However, existing simulation testing scenarios mainly rely on data-driven methods such as neural networks and reinforcement learning to mine the driver's driving habits and styles, thereby achieving background vehicle control in the testing scenario that more closely resembles human driving behavior. However, the current background vehicle trajectory generation in autonomous driving testing scenarios has at least the following shortcomings:

[0004] 1) The multi-objective trade-off mechanism is rigid and cannot flexibly cope with complex and ever-changing traffic environments; for example, when faced with complex and ever-changing traffic scenarios, it is unable to balance multiple objectives such as safety, efficiency, urgency and energy consumption in real time and accurately according to driving style.

[0005] 2) The personalized modeling of driver behavior style is crude and it is difficult to accurately simulate the diversity of human driving. If a fixed weight or simple adjustment strategy is used, it is difficult to adapt to the dynamic changes in the attention of different driving styles to each target, resulting in a lack of rationality and adaptability in the background vehicle's decision-making. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios, so as to solve the technical problems mentioned in the prior art.

[0007] A method for generating anthropomorphic trajectories for background vehicles in autonomous driving testing scenarios includes the following steps:

[0008] S1. According to the set interaction logic, extract driving features under several sets of interaction scenarios and conduct simulation training to build a driving simulator. The driving simulator uses the driving features under each set of interaction scenarios as driving behavior data of several background vehicles and configures a set driving style for each of them. At the same time, it sets a corresponding driving strategy for each driving style to build a dynamic Bayesian game framework.

[0009] S2. Real-time acquisition of the driving style of the background vehicle in any set of interactive scenarios, and iterative training of the multi-objective utility function.

[0010] S3. Calculate the payoff value corresponding to the driving style of the background vehicle using the multi-objective utility function, and select the corresponding driving strategy as the optimal response action of the background vehicle from the dynamic Bayesian game framework based on the calculated payoff value.

[0011] S4. Generate trajectory control instructions based on the selected driving strategy, and optimize the trajectory control instructions using a random perturbation mechanism to obtain the anthropomorphic trajectory of the background vehicle.

[0012] Optionally, in S1, the interaction scenario includes at least one of lane-changing game, intersection merging, and congested following.

[0013] Optionally, in S1, the driving characteristics include one or more of the following in any detection cycle: vehicle acceleration distribution, lane change frequency, following distance, steering wheel angle, and pedal travel.

[0014] Optionally, in S1, the method for setting the driving style includes:

[0015] Clustering algorithms are used to analyze the driving behavior data of each background vehicle. Through several iterations, the driving behavior data of the background vehicles are divided into corresponding clusters according to their feature similarity, and each cluster is used as a driving style.

[0016] The competitiveness level c of the corresponding driving style is quantified by the cluster center value:

[0017] c = w1·F norm +w2·D norm +w3·A norm ;

[0018]

[0019] Where c∈(0,1); F is the lane change frequency, F min For the minimum lane change frequency, F max For the maximum lane change frequency, F norm w1 represents the normalized lane change frequency, and w1 represents the feature weight of the lane change frequency; D represents the average following distance. min D is the minimum average following distance. max D is the maximum average following distance. norm w2 represents the normalized average following distance, and w2 represents the feature weight of the average following distance; A represents the acceleration variance. min To minimize the variance of acceleration, A max For the maximum acceleration variance, A norm w3 represents the normalized acceleration variance, and w3 is the feature weight of the acceleration variance.

[0020] The background vehicle is classified into the corresponding driving style category based on the competitiveness level c of the driving style.

[0021] Optionally, when calculating the competitiveness level c of driving style, considering the differences in the degree of attention drivers with different driving styles pay to driving objectives, the weight parameters corresponding to each driving feature are dynamically adjusted using an elastic network regression model. The elastic network regression model is set as follows:

[0022]

[0023] Among them, X i =[v avg a rms f LC d follow ] T v is the standardized vehicle driving feature vector. avg For the average velocity, a rms f is the root mean square acceleration; LC For the lane change frequency, d follow Following distance; β ko β represents the baseline value of the target weights in the elastic network regression model. ki ε represents the regression coefficients of the elastic network. k For the residual term, Represents the residual ε k Follows a mean of 0 and a variance of The vector follows a normal distribution; p is the number of standardized vehicle driving feature vectors.

[0024] When the elastic network regression model obtains weight combinations for different driving styles, the mapping relationship between the driving style and weight combination of the background vehicle is established as follows:

[0025]

[0026] Where N is the number of training samples, λ is the overall regularization level, α∈[0,1], and β T β1 is the regression coefficient vector of the elastic network; β2 is the L1 norm and β2 is the square of the L2 norm.

[0027] Optionally, the driving style can be set to aggressive, conservative, or balanced.

[0028] Optionally, in S1, the driving strategy of the dynamic Bayesian game framework is set as follows:

[0029] Competition strategy: Accelerate and / or change lanes in the current lane;

[0030] Concession strategy: Decelerate in the current lane or maintain a safe distance from the adjacent background vehicle;

[0031] Probing strategy: Maintain the current driving state and observe the driving state of other background vehicles in adjacent driving lanes or the same driving lane.

[0032] Optionally, in step S2, the multi-objective utility function u is set as follows:

[0033] u=ω te f te +ω sf f sf +ω ug f ug +ω ef f ef ;

[0034]

[0035] Where, ω te For efficiency weights, ω sf For safety weights, ω ug As the urgency weight, ω ef Energy consumption weights are used to simulate the trade-off mechanisms of human drivers in complex scenarios; f te For efficiency and cost, exp(·) is an exponential function operation with the natural constant e as the base, v sv Given the current speed of the background vehicle, v min v is the minimum speed allowed in the current interaction scenario. max f is the maximum speed allowed in the current interaction scenario, where ε is the smoothing factor; sf For safety costs, used to reflect the collision risk of the background vehicle, TTC is the collision time, τ is the time constant, and λ is the collision cost. d d represents the weighting coefficient. min Minimum vehicle changing distance; f ugX represents the urgency cost, used to reflect the urgency of the background vehicle's approach to the target point. int Let X be the coordinates of the starting point of the lane change along the road direction. term Y represents the coordinates of the end point of the lane change along the road direction. int Y represents the coordinates of the starting point of the lane change in the direction perpendicular to the road. term f represents the coordinates of the lane change endpoint in the direction perpendicular to the road. ef Energy cost, used to quantify the energy consumption cost of control actions, v term For the desired speed, v cur Given the current speed, a max The maximum acceleration of the background vehicle.

[0036] Optionally, in step S3, the method for selecting the appropriate driving strategy from the dynamic Bayesian game framework specifically includes:

[0037] The utility values ​​of the contention strategy and the concession strategy are calculated separately according to the multi-objective utility function to obtain the utility values ​​of the contention strategy and the concession strategy, and the relationship between the utility values ​​of the contention strategy and the concession strategy is compared.

[0038] If the utility value of the contention strategy is greater than the utility value of the concession strategy, then the contention strategy is selected;

[0039] If the utility value of the contention strategy is less than the utility value of the concession strategy, then the contention strategy is selected;

[0040] If the utility value of the competition strategy is equal to the utility value of the concession strategy, then the trial strategy is selected.

[0041] Optionally, in step S4, a trajectory control command is generated based on the current driving state of the background vehicle and the selected driving strategy. The control information of the trajectory control command includes the target speed v, acceleration a, and steering angle δ of the background vehicle; and is expressed as control parameter k. v k a k δ As influencing factors, the acceleration a and steering angle δ of the background vehicle are adjusted in real time;

[0042] a = c·k a ·a max ·sign(k v ·v target -v)+(1-c)·(v target -v)+η·rand;

[0043] δ=c·k δ ·δ max ·sign(e y )+(1-c)·e y +ηy ·rand;

[0044] Where, k a For acceleration influence factor, a max The maximum absolute value of acceleration, sign(·) is the operation sign function, and k v v is the target speed influence factor. target Let v be the target speed, v be the current speed, and k be the target speed. δ δ is the influencing factor of steering angle. max For the maximum steering angle, e y For lateral distance deviation, η and η y All are random disturbance amplitudes, η = 0.2a max η y =0.1δ max rand∈[-1,1] is a uniformly distributed random number.

[0045] The beneficial effects that this invention can produce include:

[0046] 1. This method achieves dynamic generation of multi-dimensional driving behavior: By integrating driving style clustering with a multi-objective weight dynamic adaptation mechanism, it overcomes the technical bottleneck of the single background vehicle behavior pattern in traditional methods. It can autonomously adjust the weight allocation of objectives such as safety, efficiency, and urgency according to real-time scene characteristics, generating a continuous driving style spectrum from conservative to aggressive, significantly improving the diversity and realism of background vehicle behavior in the test scenario.

[0047] 2. This method achieves accurate modeling of complex interaction logic: Based on the design of a dynamic Bayesian game framework, it realizes deep coupling between vehicle interaction strategies and competitive level parameters. Background: Vehicles can autonomously choose strategies of competition, concession, or exploration based on the surrounding environment, exhibiting human-like decision-making hesitation and strategy switching characteristics in scenarios such as intersection merging and lane games, thus solving the problem of mechanized interaction behavior in existing technologies.

[0048] 3. This method effectively improves the verifiability of anthropomorphism: An objective and quantitative anthropomorphism evaluation system is established using the Turing test method. The naturalness of behavioral patterns is verified through double-blind human-computer interaction testing, providing a more reliable test benchmark for the performance evaluation of autonomous driving systems. Attached Figure Description

[0049] Figure 1 The flowchart below shows the method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to the present invention.

[0050] Figure 2 This is a flowchart of the driving feature extraction and personalized behavior modeling process of the present invention;

[0051] Figure 3 This is a flowchart of the Bayesian game strategy generation and decision optimization process of the present invention. Detailed Implementation

[0052] 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.

[0053] Currently, with the rapid development of autonomous driving technology, testing is crucial to ensuring its safety and reliability. However, existing autonomous driving test scenarios suffer from significant shortcomings in background vehicle trajectory generation. These include: coarse personalized modeling of driver behavior styles, failing to accurately simulate the diversity of human driving; rigid multi-objective trade-off mechanisms, unable to flexibly cope with complex and changing traffic environments; and a lack of anthropomorphic evaluation methods, leading to a disconnect between simulation results and real-world driving scenarios. To address these shortcomings, [the following text is missing from the original extract]. Figures 1-3 As shown, this invention proposes a method for generating anthropomorphic trajectories of background vehicles in autonomous driving test scenarios, so as to realize the dynamic generation and verification of anthropomorphic interactive behaviors of background vehicles, which includes the following steps:

[0054] Step 1: Extract driving features from several sets of interactive scenarios according to the set interaction logic and conduct simulation training to build a driving simulator. These interactive scenarios are diverse, including but not limited to: lane-changing game scenarios, where vehicles need to find suitable opportunities to change lanes in traffic flow, involving a game of speed and distance with surrounding vehicles; intersection merging scenarios, where vehicles need to safely and efficiently merge into the main road traffic flow in complex intersections; and congested following scenarios, where vehicles maintain appropriate following distance and speed in low-speed, dense traffic flow to avoid collisions. The driving simulator uses the driving features from each set of interactive scenarios as driving behavior data for several background vehicles and configures a set driving style for each of them. Simultaneously, it sets corresponding driving strategies for each driving style to construct a dynamic Bayesian game framework.

[0055] In the above, driving characteristics include one or more of the following within any detection cycle: vehicle acceleration distribution (reflecting the frequency and force of vehicle acceleration / deceleration), lane change frequency (reflecting the activity level of the vehicle in changing driving lanes), following distance (directly related to driving safety), steering wheel angle (reflecting the vehicle's steering operation), and pedal travel (covering key information such as the operating range of the accelerator pedal and brake pedal).

[0056] In the above, the method for setting the driving style includes: using a clustering algorithm to analyze the driving behavior data of each background vehicle, and through several iterations of calculation, dividing the driving behavior data of the background vehicles into corresponding clusters according to their feature similarity, and using each cluster as a driving style; the driving style can be set as aggressive, conservative, or balanced, wherein the competitiveness level c of the corresponding driving style is quantified by the cluster center value;

[0057] c = w1·F norm +w2·D norm +w3·A norm ;

[0058]

[0059]

[0060] Where c∈(0,1); F is the lane change frequency, F min For the minimum lane change frequency, F max For the maximum lane change frequency, F norm w1 represents the normalized lane change frequency, and w1 represents the feature weight of the lane change frequency; D represents the average following distance. min D is the minimum average following distance. max D is the maximum average following distance. norm w2 represents the normalized average following distance, and w2 represents the feature weight of the average following distance; A represents the acceleration variance. min To minimize the variance of acceleration, A max For the maximum acceleration variance, A norm is the normalized acceleration variance, and w3 is the feature weight of the acceleration variance. Based on the competitive level c of driving style, background vehicles are classified into corresponding driving style categories; for example, drivers with an aggressive driving style tend to change lanes frequently and accelerate / decelerate rapidly, pursuing driving efficiency, and therefore exhibit characteristics of high lane change frequency, short following distance, and high acceleration variance; drivers with a conservative driving style prioritize driving safety, exhibiting characteristics of low lane change frequency, maintaining a large following distance, and gradual acceleration; and drivers with a balanced driving style seek a balance between efficiency and safety.

[0061] Furthermore, when calculating the competitiveness level *c* of driving style, drivers with different driving styles show varying degrees of focus on driving objectives (such as safety, efficiency, urgency, and energy consumption). Therefore, a resilient network regression model is used to dynamically adjust the weight parameters corresponding to each driving feature. The resilient network regression model is set as follows:

[0062]

[0063] Among them, Xi =[v avg a rms f LC d follow ] T v is the standardized vehicle driving feature vector. avg For the average velocity, a rms f is the root mean square acceleration, which is a characteristic quantity that measures the amplitude of acceleration fluctuations; LC For the lane change frequency, d follow Following distance; β ko β represents the baseline value of the target weights in the elastic network regression model, used to adjust the weight baseline; ki ε represents the regression coefficients of the elastic network. k For the residual term, Represents the residual ε k Follows a mean of 0 and a variance of The vector follows a normal distribution; p is the number of standardized vehicle driving feature vectors.

[0064] When the elastic network regression model obtains weight combinations for different driving styles, the mapping relationship between the driving style and weight combinations of the background vehicle is established as follows:

[0065]

[0066] Where N is the number of training samples, λ is the overall regularization level (value 0.1), α∈[0,1], β T The vector of regression coefficients for the elastic network is obtained by optimizing the objective function and is used to establish the mapping relationship between driving style features and multi-objective weights; β1 is the L1 norm and β2 is the square of the L2 norm; this facilitates the rapid matching of driving style and corresponding weight parameters based on the driving behavior data of the background vehicle.

[0067] In the above, based on the driving style set by the background vehicle, a corresponding driving strategy is configured for each category in the dynamic Bayesian game framework. Specifically, the driving strategy of the dynamic Bayesian game framework is set as follows:

[0068] Aggressive lane-changing strategy: Accelerating and / or changing lanes while in the current lane. When a vehicle in the background detects a favorable driving opportunity, such as a large gap in the lane ahead or surrounding vehicles moving slowly and affecting its own driving, it may adopt an aggressive lane-changing strategy to quickly accelerate or change lanes to gain a more advantageous driving position and improve driving efficiency.

[0069] Conservative yielding strategy: Reduce speed in the current lane or maintain a safe following distance from adjacent vehicles. When encountering potentially dangerous situations, such as a vehicle ahead suddenly slowing down or a vehicle in an adjacent lane attempting to change lanes abruptly, the vehicle in the background will adopt a yielding strategy to avoid collisions, reducing speed and increasing following distance to ensure driving safety.

[0070] Probing strategy (equilibrium type): Maintain the current driving state and observe the driving state of other background vehicles in adjacent lanes or the same lane. In some uncertain traffic situations, such as unclear road conditions ahead or unpredictable behavior of surrounding vehicles, background vehicles will first adopt a probing strategy, maintain the current state, gather more information, and then make a decision.

[0071] Step 2: Real-time acquisition of the driving style of background vehicles in any set of interactive scenarios, and iterative training of the multi-objective utility function based on this. The multi-objective utility function comprehensively considers multiple key dimensions such as efficiency, safety, urgency, and energy consumption, aiming to simulate the trade-off mechanisms of human drivers in complex scenarios, making the decisions of the background vehicles more closely resemble real human driving behavior. Specifically, the multi-objective utility function u is set as follows:

[0072] u=ω te f te +ω sf f sf +ω ug f ug +ω ef f ef ;

[0073]

[0074] Where, ω te For efficiency weights, ω sf For safety weights, ω ug As the urgency weight, ω ef Energy consumption weights are used to simulate the trade-off mechanisms of human drivers in complex scenarios; f te For efficiency and cost, exp(·) is an exponential function operation with the natural constant e (approximately 2.71828) as its base, v sv Given the current speed of the background vehicle, v min v is the minimum speed allowed in the current interaction scenario. max f is the maximum speed allowed in the current interaction scenario, where ε is the smoothing factor; sf For safety costs, used to reflect the collision risk of the background vehicle, TTC is the time to collision, which is the time required from the current moment for the vehicle and the obstacle to collide while maintaining the current speed and acceleration; τ is the time constant, λ dd is a weighting coefficient; the larger its value, the more significant its impact on safety costs, indicating a stricter safety constraint on minimum distance. min Minimum vehicle changing distance; f ug X represents the urgency cost, used to reflect the urgency of the background vehicle's approach to the target point. int Let X be the coordinates of the starting point of the lane change along the road direction. term Y represents the coordinates of the end point of the lane change along the road direction. int Y represents the coordinates of the starting point of the lane change in the direction perpendicular to the road. term f represents the coordinates of the lane change endpoint in the direction perpendicular to the road. ef Energy cost, used to quantify the energy consumption cost of control actions, v term For the desired speed, v cur Given the current speed, a max The maximum acceleration of the background vehicle is given. During iterative training, these parameters are continuously optimized and adjusted to ensure the multi-objective utility function can adapt to different driving styles, enabling the background vehicle to make reasonable decisions in various scenarios and enhancing the flexibility and adaptability of the strategy.

[0075] Step 3: Calculate the payoff value corresponding to the driving style of the background vehicle using a multi-objective utility function, and select the corresponding driving strategy as the optimal response action of the background vehicle from the dynamic Bayesian game framework based on the calculated payoff value. Specifically, this includes: calculating the utility values ​​of the competing strategy and the conceding strategy respectively according to the multi-objective utility function, obtaining the utility value u(A) of the competing strategy. F ) and the utility value of the concession strategy u(A) Y ), and compare the utility value u(A) of the competition strategy. F ) and the utility value of the concession strategy u(A) Y The relationship between the magnitudes of the strategies (A, B, C, D, E) is as follows; where, if the utility value of the competing strategy is u(A, B, C, D, E), then... F The utility value of the concession strategy is greater than u(A). Y If the utility value of the competition strategy is u(A), it indicates that under the current circumstances, adopting the competition strategy can enable the background vehicle to obtain higher overall benefits, such as reaching the destination faster and utilizing road resources more efficiently. Therefore, the competition strategy is chosen. F The utility value of the concession strategy is less than u(A). Y If the yield value is less than the required value, it indicates that the concession strategy better meets the needs of the current scenario, can better ensure driving safety and avoid potential dangers, and the concession strategy should be chosen in this case; if ... F The utility value of the concession strategy is equal to u(A). Y If the overall benefits of the two strategies are roughly equal under the current circumstances, it is difficult to directly judge which is better. In this case, the trial strategy is chosen to maintain the current driving state and further observe the surrounding traffic conditions, so as to provide a more sufficient basis for subsequent decision-making.

[0076] Step 4: Generate trajectory control commands based on the selected driving strategy, and optimize these commands using a random perturbation mechanism to obtain a human-like trajectory for the background vehicle. Specifically, based on the current driving state of the background vehicle and the selected driving strategy, trajectory control commands are generated. The control information in these commands includes the target speed v, acceleration a, and steering angle δ of the background vehicle. To make the generated trajectory more human-like, a random perturbation mechanism is introduced to control parameter k. v k a k δ The background vehicle's acceleration 'a' and steering angle 'δ' are adjusted in real time as influencing factors; the trajectory control command is optimized in a human-like manner to obtain the final human-like trajectory.

[0077] a = c·k a ·a max ·sign(k v ·v target -v)+(1-c)·(v target -v)+η·rand;

[0078] δ=c·k δ ·δ max ·sign(e y )+(1-c)·e y +η y ·rand;

[0079] Where, k a For acceleration influence factor, a max The maximum absolute value of acceleration; sign(·) is the operation sign function, which returns 1 for a positive input, 0 for a negative input, and -1 for a negative input; k v v is the target speed influence factor. target Let v be the target speed, v be the current speed, and k be the target speed. δ δ is the influencing factor of steering angle. max For the maximum steering angle, e y For lateral distance deviation, η and η y All are random disturbance amplitudes, η = 0.2a max η y =0.1δ max rand∈[-1,1] is a uniformly distributed random number used to simulate the uncertainty of the driver's throttle and steering wheel control.

[0080] In this embodiment, different driving strategies correspond to different control parameter settings, as detailed below:

[0081] (1) Control parameter k of the contention strategy v k ak δ Set them to:

[0082] k v =1+0.5c;k a =1+0.6c;k δ =1 + 0.6c;

[0083] (2) Control parameter k of the concession strategy v k a k δ Set them to:

[0084] k v =1-0.3c;k a =1-0.4c; k δ =1-0.4c;

[0085] (3) Control parameter k of the trial strategy v k a k δ All are set to 1.

[0086] Step 5: To scientifically and accurately verify the anthropomorphism of the background vehicle's driving behavior data, this invention employs the classic Turing test method to verify the anthropomorphism level of the final obtained background vehicle's trajectory. Specifically, experienced testers are recruited to interact with the background vehicle in real-time in typical interactive scenarios (such as lane-changing maneuvers, intersection merging, and congested following). It should be noted that testers need to rely on their own driving experience and intuition to judge whether the background vehicle is under human control and evaluate its driving style (aggressive / conservative / balanced). Based on the test results, the differences between the background vehicle's driving behavior data and real human driving behavior are analyzed in depth, and the driving simulator is iteratively optimized. By adjusting algorithm parameters, such as weight parameters in the elastic network regression model, weight coefficients in the multi-objective utility function, and optimizing the strategy selection mechanism, the driving behavior of the background vehicle in complex traffic scenarios becomes more realistic and natural, gradually narrowing the gap between simulation results and real-world scenarios. This ensures that the driving behavior and trajectory of the background vehicle in the simulation test scenario closely resemble real-world conditions, providing more reliable and effective data support for autonomous driving testing.

[0087] Example 1

[0088] This invention uses Carla autonomous driving simulation software to generate and test vehicle trajectories.

[0089] Before starting the simulation environment, a high-precision driving simulator was used to collect driving behavior data from human drivers. Specifically, 50 experienced drivers were recruited to participate in the simulated driving experiment. Each driver was required to complete 10 repeated tests in typical scenarios such as lane-changing maneuvering, intersection merging, and congested following. The simulator recorded key feature data such as speed sequence, acceleration distribution, lane-changing frequency, and following distance at a sampling frequency of 100Hz. The core features in the driving behavior data were calculated: acceleration variance A, average following distance D, and lane-changing frequency F. These three core features were then normalized to eliminate differences in units and unify the direction to positive (i.e., the larger the value, the stronger the competitiveness). Then, an improved K-means clustering algorithm (cluster number = 3, iterations 100 times) was used to classify driving styles, ultimately dividing the driving behavior data into three driving styles:

[0090] Aggressive type (32%): F norm ≥0.7, D norm ≥0.7, A norm ≥0.65, c≥0.68;

[0091] Conservative type (41%): F norm ≤0.3, D norm ≤0.3, A norm ≤0.35, c≤0.32;

[0092] Balanced type (accounting for 27%): The eigenvalues ​​of acceleration variance A, average following distance D, and lane change frequency F are in the middle distribution, with 0.32 < c < 0.68.

[0093] Based on the clustering results, an elastic network regression model is used to dynamically generate weight parameters for the multi-objective utility functions corresponding to the three driving styles. The specific weight range is as follows:

[0094] Aggressive type (c≥0.68): Efficiency weight ω te =0.68, safety weight ω sf =0.12, urgency weight ω ug =0.09, energy consumption weight ω ef =0.11;

[0095] Conservative type (c≤0.32): Efficiency weight ω te =0.18, safety weight ω sf =0.72, urgency weight ω ug =0.06, energy consumption weight ω ef =0.04;

[0096] Balanced type (0.32 < c < 0.68): Efficiency weight ω te =0.25, safety weight ωsf =0.25, urgency weight ω ug =0.25, energy consumption weight ω ef =0.25.

[0097] Once the simulation environment starts, an initial c-value is first assigned to each background vehicle. Then, a baseline value is set according to the vehicle type and dynamically corrected based on the initial position. Specifically, the baseline value c for sedans is... base =0.5, truck baseline value c base =0.3, sports car c base =0.7; vehicles in the first 20% of the traffic flow c init =c base +0.1rand, vehicles in the bottom 20% of the list. init =c base -0.1rand, the remaining vehicles c init =c base +0.05rand. Where rand∈[-0.05, 0.05] is a uniformly distributed random number. For example, in the first 20% of a sports car in the traffic flow, the initial competition value c init =0.7 + 0.1 × 0.3 = 0.73, which belongs to the aggressive type; a truck in the last 20% of the traffic flow, the initial competition value c init =0.3-0.1×0.2=0.28, which belongs to the conservative type.

[0098] During the driving process, the background vehicle dynamically adjusts its competitiveness level (c) based on real-time driving data. For example, in a lane-changing scenario, successfully seizing the target lane (c)... new =c current +0.05×(1-c current Forced to abandon lane change c new =c current -0.03×c current The background vehicle can match its driving style and corresponding multi-objective utility function weight parameters based on the current competitiveness c. During the journey, the background vehicle makes decisions using a dynamic Bayesian game framework. For example, when background vehicle A (aggressive, c = 0.82) and the autonomous vehicle B are driving side-by-side on a highway, and background vehicle A detects that the autonomous vehicle B intends to change lanes, the utility values ​​of each strategy are calculated as follows: Competition strategy u(A) F =0.76, concession strategy u(A) Y If ) = 0.34, then background vehicle A chooses a competition strategy and accelerates to v. target =k v ·v current =1.41v current Ultimately, they successfully prevented the autonomous vehicle under test, B, from changing lanes.

[0099] Anthropomorphic Verification Phase: Thirty professional drivers (≥5 years of driving experience) were recruited. Each driver interacted in real-time with a total of 100 background vehicles (32 aggressive, 41 conservative, and 27 balanced) in three typical scenarios: lane changing on highways, merging at intersections without traffic lights, and following other vehicles in congested traffic. Each test lasted 3 minutes. Test results showed that the average percentage of background vehicles misidentified as human drivers (behavioral confusion rate) reached 79.9%, with the highest confusion rate among aggressive vehicles (83.5%), followed by conservative vehicles at 76.8%, and balanced vehicles at 79.3%. These data indicate that the anthropomorphic trajectory interaction behavior of the background vehicles generated by the algorithm closely resembles real driving habits, verifying the effectiveness and reliability of the method.

[0100] Example 2

[0101] The difference between this embodiment and Embodiment 1 is that the background vehicle trajectory is generated and tested using Unity autonomous driving simulation software. Furthermore, depending on different testing requirements, when the background vehicle connects to V2X (Vehicle-to-Everything) communication, its c-value change trend can be broadcast to the vehicle under test in real time. This function is of great significance in practical testing. For example, in complex traffic intersection scenarios, the driving intention of the background vehicle may change rapidly with changes in the surrounding environment. By sending the c-value change trend to the vehicle under test, the system under test can be tested for its responsiveness to changes in the background vehicle's intentions, evaluating the accuracy and timeliness of the autonomous driving system's decision-making in dynamic traffic environments, and meeting the diverse needs for performance evaluation of autonomous driving systems in different testing scenarios.

Claims

1. A method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios, characterized in that, Includes the following steps: S1. A driving simulator is constructed by extracting driving features from several sets of interactive scenarios according to the set interaction logic and performing simulation training. The driving simulator uses the driving features from each set of interactive scenarios as driving behavior data for several background vehicles and configures a set driving style for each of them. Simultaneously, a corresponding driving strategy is set for each driving style to construct a dynamic Bayesian game framework. The driving strategy of the dynamic Bayesian game framework is set as follows: Competition strategy: Accelerate and / or change lanes in the current lane; Concession strategy: Decelerate in the current lane or maintain a safe distance from the adjacent background vehicle; Probing strategy: Maintain the current driving state and observe the driving state of other background vehicles in adjacent driving lanes or the same driving lane; S2. Real-time acquisition of the driving style of the background vehicle in any set of interactive scenarios, and iterative training of the multi-objective utility function, wherein the multi-objective utility function u is set as follows: ； ； ； ； ； in, For efficiency weighting, For safety weights, As a weight for urgency, Energy consumption weights are used to simulate the trade-off mechanisms of human drivers in complex scenarios. For efficiency and cost, This refers to the operation of an exponential function with the natural constant e as the base. The current speed of the background vehicle. The minimum speed allowed in the current interaction scenario. The maximum speed allowed in the current interaction scenario. It is a smoothing factor; For safety costs, used to reflect the collision risk of the background vehicle; TTC stands for Time to Collision. It is a time constant. These are the weighting coefficients. Minimum vehicle changing distance; The urgency cost is used to reflect the urgency of the background vehicle's approach to the target point. The coordinates of the starting point of the lane change along the road direction. The coordinates are the endpoints of the lane change along the road direction. The coordinates of the starting point of the lane change in the direction perpendicular to the road. The coordinates of the end point of the lane change in the direction perpendicular to the road; Energy cost, used to quantify the energy consumption of control actions. For the desired speed, At the current speed, The maximum acceleration of the background vehicle; S3. Calculate the payoff value corresponding to the driving style of the background vehicle using the multi-objective utility function, and select the corresponding driving strategy as the optimal response action of the background vehicle from the dynamic Bayesian game framework based on the calculated payoff value. S4. Generate trajectory control instructions based on the selected driving strategy, and optimize the trajectory control instructions using a random perturbation mechanism to obtain the anthropomorphic trajectory of the background vehicle.

2. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 1, characterized in that, In S1, the interaction scenario includes at least one of lane-changing game, intersection merging, and congested following.

3. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 1, characterized in that, In S1, the driving characteristics include one or more of the following in any detection cycle: vehicle acceleration distribution, lane change frequency, following distance, steering wheel angle, and pedal travel.

4. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 1, characterized in that, In S1, the method for setting the driving style includes: Clustering algorithms are used to analyze the driving behavior data of each background vehicle. Through several iterations, the driving behavior data of the background vehicles are divided into corresponding clusters according to their feature similarity, and each cluster is used as a driving style. The competitiveness level c of the corresponding driving style is quantified by the cluster center value: ； ； ； ； in, F is the channel switching frequency, F min For the minimum lane change frequency, F max For the maximum lane change frequency, F norm w1 represents the normalized lane change frequency, and w1 represents the feature weight of the lane change frequency; D represents the average following distance. min D is the minimum average following distance. max D is the maximum average following distance. norm w2 represents the normalized average following distance, and w2 represents the feature weight of the average following distance; A represents the acceleration variance. min To minimize the variance of acceleration, A max For the maximum acceleration variance, A norm w3 represents the normalized acceleration variance, and w3 is the feature weight of the acceleration variance. The background vehicle is classified into the corresponding driving style category based on the competitiveness level c of the driving style.

5. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 4, characterized in that, When calculating the competitiveness level *c* of driving style, considering the differences in attention drivers pay to driving objectives based on different driving styles, a resilient network regression model is used to dynamically adjust the weight parameters corresponding to each driving feature. The resilient network regression model is set as follows: ； in, This is the standardized vehicle driving feature vector. For average speed, It is the root mean square acceleration; For lane change frequency, Following distance; These are the base values ​​for the target weights in the elastic network regression model; These are the regression coefficients of the elastic network; For the residual term, Residual Follows a mean of 0 and a variance of The vector follows a normal distribution; p is the number of standardized vehicle driving feature vectors. When the elastic network regression model obtains weight combinations for different driving styles, the mapping relationship between the driving style and weight combination of the background vehicle is established as follows: ； Where N is the number of training samples. For the overall regularization level, , This is the vector of regression coefficients for the elastic network; It is an L1 norm. It is the square of the L2 norm.

6. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 1, characterized in that, The driving styles are set as aggressive, conservative, and balanced.

7. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 1, characterized in that, In S3, the method for selecting the appropriate driving strategy from the dynamic Bayesian game framework specifically includes: The utility values ​​of the contention strategy and the concession strategy are calculated separately according to the multi-objective utility function to obtain the utility values ​​of the contention strategy and the concession strategy, and the relationship between the utility values ​​of the contention strategy and the concession strategy is compared. If the utility value of the contention strategy is greater than the utility value of the concession strategy, then the contention strategy is selected; If the utility value of the contention strategy is less than the utility value of the concession strategy, then the contention strategy is selected; If the utility value of the competition strategy is equal to the utility value of the concession strategy, then the trial strategy is selected.

8. The method for generating anthropomorphic trajectories of background vehicles for autonomous driving test scenarios according to claim 1, characterized in that, In step S4, a trajectory control command is generated based on the current driving state of the background vehicle and the selected driving strategy. The control information of the trajectory control command includes the target speed v and acceleration of the background vehicle. and steering angle ; and with control parameters , , As an influencing factor, the acceleration of the background vehicle is adjusted in real time. and steering angle ; ； ； in, As the acceleration influencing factor, The absolute value of the maximum acceleration. For operations, refer to the sign function. As the target speed influencing factor, Let v be the target speed and v be the current speed. The influencing factor is the steering angle. For the maximum steering angle, This refers to the lateral distance deviation. and All are random disturbance amplitudes. =0.2 , =0.1 rand These are evenly distributed random numbers.

Images