Scene generation method and system for automatic driving cold chain logistics vehicle test evaluation

Abstract

The application discloses a kind of scene generation method and system for automatic driving cold chain logistics vehicle test evaluation, which comprises the following steps: collecting the multidimensional data set of cold chain logistics vehicle driving scene, filtering lane-changing cut-in scene;According to the head time interval, longitudinal correction collision time and transverse collision time, the comprehensive cognitive risk score of each lane-changing cut-in scene is calculated, the risk level of lane-changing cut-in scene is obtained based on clustering classification, and the trajectory sequence of high-risk lane-changing cut-in scene is extracted;The physical model of cold chain logistics vehicle is constructed, the acceleration of cold chain logistics vehicle is calculated, and the driving trajectory of cold chain logistics vehicle and target vehicle lane-changing scene is generated;The conditional vector and the trajectory sequence of high-risk lane-changing cut-in scene are input into conditional variational autoencoder to obtain the reconstructed trajectory sequence;An automatic driving test scene library based on the characteristics of cold chain logistics vehicle is constructed.The application can realize the accurate generation of high-coverage automatic driving test scene for the characteristics of cold chain logistics vehicle.

Description

Technical Field

[0001] This invention relates to the field of vehicle test scenario simulation technology, specifically to a scenario generation method and system for testing and evaluating autonomous cold chain logistics vehicles. Background Technology

[0002] The practical application of autonomous driving technology requires thorough scenario testing and verification. However, as a special commercial vehicle, the characteristics of cold chain logistics vehicles, such as load variation, refrigeration system operation, and low-temperature environment adaptability, make the test scenario generation methods of ordinary passenger cars unsuitable for direct application.

[0003] Existing methods for generating test scenarios for ordinary passenger vehicles have three major flaws: First, data collection does not focus on cold chain characteristics, lacking the integration of key parameters such as load, refrigeration power, and cargo compartment temperature and humidity with traffic behavior data; second, risk assessment only focuses on longitudinal safety indicators (such as THW and MTTC), ignoring lateral collision risks (such as TTClat), and does not combine the dynamic characteristics of cold chain logistics vehicles to quantify risk levels; third, the scenario generation model does not incorporate the physical constraints of cold chain logistics vehicles and their environmental coupling, resulting in deviations in the physical feasibility, safety, and statistical authenticity of generated lane-changing and other scenario trajectories. These characteristics make the safety and reliability testing of cold chain logistics vehicles in complex scenarios face higher requirements.

[0004] Specifically, existing test scenario generation methods are mainly based on physical models and data-driven methods, but they face significant challenges when adapting to cold chain logistics vehicles:

[0005] Physical model-based methods construct vehicle dynamics models based on physical principles and laws. However, they struggle to fully consider the various complex factors and uncertainties in actual operation, failing to model the complex coupling relationships between the refrigeration system, external environment, and vehicle power. Furthermore, in practical applications, it is difficult to accurately obtain parameters such as road rolling resistance coefficient and wind resistance coefficient, affecting model accuracy. Secondly, these methods make many idealized assumptions about the vehicle system and environment, neglecting some real-world complexities, such as the impact of load and refrigeration system power on vehicle dynamics (driving force). This leads to significant deviations in vehicle acceleration prediction, resulting in insufficient adaptability to complex real-world scenarios and making them unsuitable for scenario testing of cold chain logistics vehicles. Additionally, the model construction and calculation process is relatively complex, requiring high levels of professional knowledge and computational resources, hindering rapid promotion and application.

[0006] Data-driven methods establish linear or nonlinear functional relationships between the vehicle's acceleration and variables such as distance, position, speed difference, and collision time between the vehicle and the vehicle in front, based on information obtained from onboard sensors. This enables prediction of vehicle trajectories in driving scenarios. However, due to limitations in the sensing range and sample coverage of onboard sensors, it is difficult to effectively integrate macroscopic traffic data (such as weather, road conditions, and multi-vehicle interaction patterns), resulting in weak generalization ability and low prediction accuracy in extreme conditions. Furthermore, natural driving data for cold chain logistics vehicles is scarce, especially lacking their unique environmental parameters and vehicle operating status, leading to insufficient reliability of test results. In addition, the deterministic relationship models established by linear or nonlinear functions between the vehicle's acceleration and factors such as distance and speed difference between the vehicle and the vehicle in front fail to fully consider the differences in driving behavior among different drivers and vehicle dynamic constraints, making it difficult to reflect the diversity of real-world scenarios.

[0007] In summary, there are problems such as insufficient coverage, unreasonable generated scenarios, and low generation efficiency in the generation of test scenarios for autonomous driving cold chain logistics vehicles, which make it difficult to meet the needs of real-time optimization. Therefore, there is an urgent need for a technology that can quickly generate high-coverage autonomous driving test scenarios tailored to the characteristics of cold chain logistics vehicles in order to improve their operational efficiency in complex environments. Summary of the Invention

[0008] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a scenario generation method and system for testing and evaluating autonomous driving cold chain logistics vehicles. Starting with a multidimensional dataset of driving scenarios for cold chain logistics vehicles, this invention calculates a comprehensive cognitive risk score for each lane-changing scenario based on the headway, longitudinal corrected collision time, and lateral collision time. It then extracts trajectory sequences of high-risk lane-changing scenarios using a comprehensive longitudinal and lateral cognitive risk assessment method. A physical model of the cold chain logistics vehicle is constructed based on its dynamic characteristics. The acceleration of the cold chain logistics vehicle is calculated using the longitudinal dynamic equation, generating the driving trajectory of the cold chain logistics vehicle in lane-changing scenarios with the target vehicle. Finally, the invention integrates conditional vectors with high-risk lane-changing scenarios. The trajectory sequence input conditional variational autoencoder for the entry scenario generates a reconstructed trajectory sequence, constructing an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles. While efficiently generating a rich variety of test scenarios, it conducts a comprehensive analysis of the generated scenarios, including distribution similarity, physical feasibility, and realism, forming a closed-loop logic for test generation and evaluation, encompassing risk data extraction, dynamic model construction, scenario generation, and assessment. This can efficiently generate test cases covering high-risk scenarios such as extreme weather and sudden congestion, reducing parameter dimensionality, significantly shortening the test cycle, and providing a highly reliable decision-making basis for autonomous driving systems. This solves the problems of low coverage and insufficient generation efficiency of test scenarios for autonomous driving cold chain logistics vehicles.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention provides a scenario generation method for testing and evaluating autonomous cold chain logistics vehicles, comprising the following steps:

[0011] Collect a multidimensional dataset of driving scenarios for cold chain logistics vehicles, and filter lane-changing entry scenarios based on preset scenario filtering rules;

[0012] The comprehensive cognitive risk score for each lane change scenario is calculated based on the headway, longitudinal correction collision time, and lateral collision time. All lane change scenarios are clustered and classified to obtain the risk level of each lane change scenario. The trajectory sequence of high-risk lane change scenarios is extracted.

[0013] A physical model of the cold chain logistics vehicle is constructed based on its dynamic characteristics, and the longitudinal dynamic equation of the cold chain logistics vehicle is obtained. The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation, and the driving trajectory of the cold chain logistics vehicle in the lane-changing scenario with the target vehicle is generated.

[0014] Construct a conditional vector, input the conditional vector and the trajectory sequence of the high-risk lane change scenario into a conditional variational autoencoder to obtain a reconstructed trajectory sequence, and decouple to obtain motion feature parameters;

[0015] The differences in motion characteristic parameter distribution between the reconstructed trajectory sequence and real high-risk lane change and cut-in scenarios were evaluated, and an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles was constructed.

[0016] As a preferred technical solution, the multidimensional dataset for cold chain logistics vehicle driving scenarios includes cold chain logistics vehicle operation data, meteorological environment data, and driving behavior data.

[0017] As a preferred technical solution, the preset scenario selection rule is: the cold chain logistics vehicle is the main vehicle, the target lane for the vehicle in front to change lanes is the lane where the cold chain logistics vehicle is located, and after changing lanes, it is in front of the cold chain logistics vehicle, and there are no other vehicles between the two vehicles.

[0018] As a preferred technical solution, the acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation, and expressed as:

[0019] ;

[0020] in, This indicates the acceleration of the cold chain logistics vehicle. This represents the road adhesion coefficient of a dry road surface. It is the attenuation coefficient of the road surface adhesion coefficient with precipitation. Indicates precipitation. It is the acceleration due to gravity. This is the power loss coefficient of cooling capacity. For the power consumed by the cooling system, This indicates the current speed of the cold chain logistics vehicle. For the total mass of the cold chain logistics vehicle, Indicates the rolling resistance coefficient. This refers to the longitudinal wind resistance experienced by the vehicle.

[0021] As a preferred technical solution, the driving trajectory of the cold chain logistics vehicle and the target vehicle in the lane-changing scenario is generated, specifically including:

[0022] The minimum safe distance is calculated based on the distance traveled during constant speed driving, variable deceleration driving, and constant deceleration driving, as well as the length of the vehicle changing lanes. The longitudinal position, lateral position, and initial speed of the cold chain logistics vehicle in the initial state are then determined.

[0023] The cold chain logistics vehicle starts from the initial state and decelerates at maximum speed until its speed drops to the same speed as the target vehicle at the moment of lane change. The cold chain logistics vehicle then generates a following trajectory based on the target vehicle's trajectory.

[0024] As a preferred technical solution, constructing a condition vector specifically includes:

[0025] A condition vector is constructed by selecting the initial longitudinal velocity, initial lateral velocity, lateral displacement, peak lateral velocity, the time frame corresponding to the peak lateral velocity, and the road adhesion coefficient.

[0026] As a preferred technical solution, the conditional vector and the trajectory sequence of the high-risk lane-change scenario are input into a conditional variational autoencoder to obtain a reconstructed trajectory sequence, specifically including:

[0027] The conditional vector and the trajectory sequence of the high-risk lane-changing cutting scenario are encoded by the posterior distribution parameters of the latent space, and the latent vector is randomly sampled based on the posterior distribution parameters and noise.

[0028] The latent vector and the conditional vector are concatenated and input into the decoder, and the decoded output is a reconstructed trajectory sequence.

[0029] As a preferred technical solution, in the step of reconstructing the trajectory sequence, a total loss function is constructed based on trajectory reconstruction loss, KL divergence loss, distribution constraint loss, smoothness loss, and physical consistency loss;

[0030] The trajectory reconstruction loss is calculated using the mean squared error to determine the difference between the reconstructed trajectory sequence and the true trajectory.

[0031] KL divergence loss calculates the similarity between the posterior distribution of a latent variable and the standard normal prior.

[0032] The distribution constraint loss, based on one-dimensional Wasserstein distance constraints, reconstructs the motion feature statistical distribution of the trajectory sequence, which is consistent with the true trajectory.

[0033] The smoothness loss uses L2 loss constraints to reconstruct the magnitude of the lateral velocity variation of the trajectory sequence in the first difference over time.

[0034] The change in lateral displacement of the reconstructed trajectory sequence using L2 loss constraint is approximately equal to the discrete integral of the lateral velocity.

[0035] As a preferred technical solution, the step of reconstructing the trajectory sequence further includes a trajectory adjustment and fitting step, specifically including:

[0036] The velocity variation component of the reconstructed trajectory sequence is adjusted proportionally based on the road surface adhesion coefficient.

[0037] Polynomial fitting is performed on the reconstructed trajectory sequence.

[0038] The present invention also provides a scenario generation system for testing and evaluating autonomous driving cold chain logistics vehicles, which is used to execute the above-mentioned scenario generation method for testing and evaluating autonomous driving cold chain logistics vehicles, including: a data acquisition module, a scenario screening module, a risk assessment module, a trajectory extraction module, a driving trajectory generation module, a trajectory sequence reconstruction module, and a scenario library construction module;

[0039] The data acquisition module is used to collect multidimensional datasets of cold chain logistics vehicle driving scenarios;

[0040] The scene filtering module is used to filter lane-change entry scenes based on preset scene filtering rules;

[0041] The risk assessment module is used to calculate the comprehensive cognitive risk score of each lane change scenario based on the headway, longitudinal correction collision time, and lateral collision time, and to cluster and classify all lane change scenarios to obtain the risk level of the lane change scenario.

[0042] The trajectory extraction module is used to extract trajectory sequences for high-risk lane-changing and cutting-in scenarios;

[0043] The driving trajectory generation module is used to generate the driving trajectory of the cold chain logistics vehicle in the lane-changing scenario with the target vehicle, specifically including:

[0044] A physical model of the cold chain logistics vehicle is constructed based on its dynamic characteristics, and the longitudinal dynamic equation of the cold chain logistics vehicle is obtained. The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation, and the driving trajectory of the cold chain logistics vehicle in the lane-changing scenario with the target vehicle is generated.

[0045] The trajectory sequence reconstruction module is used to input the conditional vector and the trajectory sequence of the high-risk lane change scene into the conditional variational autoencoder to obtain the reconstructed trajectory sequence and decouple to obtain motion feature parameters.

[0046] The scenario library construction module is used to evaluate the differences in motion feature parameter distribution between the reconstructed trajectory sequence and the real high-risk lane change and cutting scenario, and to construct an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] (1) This invention provides data support for the accurate generation of trajectories in high-coverage autonomous driving test scenarios oriented towards the characteristics of cold chain logistics vehicles by collecting datasets with characteristics of cold chain logistics vehicles. The comprehensive cognitive risk score of each lane change scenario is calculated based on the headway, longitudinal correction collision time, and lateral collision time. The risk level of the scenario is quantitatively characterized by the comprehensive cognitive risk assessment method of longitudinal and lateral, avoiding the limitations of a single indicator and making the risk assessment more comprehensive and accurate.

[0049] (2) This invention constructs a physical model of a cold chain logistics vehicle based on its dynamic characteristics, which can more realistically reflect the driving characteristics of the cold chain logistics vehicle. The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation. Furthermore, when calculating the acceleration of the cold chain logistics vehicle, key factors such as the dynamic influence of load change on vehicle inertial parameters and the nonlinear effect of weather on road surface adhesion coefficient are embedded into the dynamic model to construct a two-dimensional physical model that couples dynamic characteristics with the environment. The influence of cold chain characteristics on vehicle response is quantified, and the longitudinal dynamic equation of the cold chain logistics vehicle containing the influence of load, refrigeration power and precipitation is obtained. This solves the problem of vehicle dynamic modeling deviation caused by the complexity of parameter coupling and insufficient environmental adaptability of traditional methods. In turn, the driving trajectory of the cold chain logistics vehicle and the target vehicle in the lane-changing scenario is generated, realizing the high-precision generation of vehicle behavior trajectory of the cold chain logistics vehicle under complex working conditions such as load change and extreme weather.

[0050] (3) In this invention, the conditional vector and the trajectory sequence of the high-risk lane-changing entry scenario are input into the conditional variational autoencoder to obtain the reconstructed trajectory sequence. The conditional variational autoencoder learns the feature distribution of the high-risk lane-changing trajectory and generates a trajectory that conforms to the real distribution. Furthermore, the reconstructed trajectory sequence is driven by physical constraints, and constraints such as distribution similarity, dynamic feasibility and trajectory smoothness are introduced. This can generate lane-changing trajectories that conform to both real driving behavior and vehicle dynamics, effectively improving the diversity, smoothness and physical feasibility of trajectory generation. Through the deep integration of the data distribution learning of the generation model and the prior constraints of the physical model, the high-coverage autonomous driving test scenario for cold chain logistics vehicles can be accurately generated. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to the present invention.

[0052] Figure 2This is a schematic diagram of the clustering distribution results of lane-changing entry scenarios with different risk levels according to the present invention.

[0053] Figure 3 This is a schematic diagram showing the kinematic parameter feature distribution of lane change scenarios with different risk levels according to the present invention.

[0054] Figure 4 This is a schematic diagram of the following driving trajectory of the cold chain logistics vehicle of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] Example 1

[0057] like Figure 1 As shown, this embodiment provides a scenario generation method for testing and evaluating autonomous driving cold chain logistics vehicles. Due to the characteristics of cold chain logistics vehicles, such as cargo temperature control requirements, variable loads, and fixed routes, directly applying scenario generation methods used for passenger cars to cold chain logistics vehicles results in scenarios that cannot meet the actual testing needs of these vehicles, leading to insufficient reliability of test results. This embodiment proposes a comprehensive longitudinal and lateral cognitive risk assessment method based on onboard sensor data, environmental perception data, and roadside sensor data from the cold chain logistics vehicle. It extracts risk scenario data to provide a real data foundation for subsequent model training; and constructs a dynamic characteristic-environment coupling characteristic... A multi-dimensional parametric vehicle dynamics characterization model is used to generate autonomous vehicle driving trajectories with cold chain logistics vehicle characteristics. A data-driven scenario generation model integrating physical constraints and a multi-dimensional evaluation mechanism are established to efficiently generate diverse test scenarios. Simultaneously, the generated scenarios are analyzed from multiple dimensions, including distribution similarity, physical feasibility, and realism / human-likeness. This provides cold chain logistics vehicles with richer, more realistic, and more comprehensive driving scenarios, improving the testing and evaluation capabilities of autonomous cold chain logistics vehicles and solving problems such as insufficient coverage, low generation efficiency, and a single evaluation method in the generation of test scenarios for autonomous cold chain logistics vehicles. Specifically, the steps include:

[0058] S1: Collect a multidimensional dataset of driving scenarios for cold chain logistics vehicles;

[0059] In this embodiment, the multidimensional dataset of the cold chain logistics vehicle driving scenario includes cold chain logistics vehicle operation data, meteorological environment data, and driving behavior data;

[0060] Specifically, the vehicle-mounted sensors collect operating data of the cold chain logistics vehicle, including data such as the temperature of the compartment, the weight of the goods, and the power of the refrigeration system.

[0061] Meteorological and environmental data, including temperature, relative humidity, rainfall, and snowfall, are collected through vehicle-mounted weather stations.

[0062] Information such as vehicle type and road type is recorded through roadside sensors;

[0063] The system collects driving behavior data of the vehicle and surrounding vehicles through in-vehicle cameras, millimeter-wave radar, and GPS modules, including information such as vehicle speed, acceleration, heading angle, distance between vehicles, collision time, lane ID, and vehicle position.

[0064] In this embodiment, the collected raw data is stored and managed in a standardized manner according to the consistency principle through a timestamp synchronization mechanism, forming a multidimensional dataset of cold chain logistics vehicle driving scenarios;

[0065] S2: Filter lane-changing entry scenarios based on preset scenario filtering rules;

[0066] In this embodiment, the preset scenario filtering rule is: the cold chain logistics vehicle is the main vehicle, the target lane of the vehicle in front changing lanes is the lane where the cold chain logistics vehicle is located, and after changing lanes, it is in front of the cold chain logistics vehicle, and there are no other vehicles between the two vehicles.

[0067] Specifically, in the lane-changing scenario, the cold chain logistics vehicle does not change lanes and is in a stable following state. The target vehicle in front has changes in lateral acceleration, lateral position, and heading angle over a set time, and the relative lateral and longitudinal distance between the vehicle and the vehicle in front decreases.

[0068] This embodiment uses the highway natural driving dataset from the high D dataset as an example. It filters out target vehicle lane-changing scenarios from the dataset. A change in lane ID is considered a lane change process. The trajectory data of the driver and target vehicle for 50 frames before and after the lane change are extracted. That is, the driving trajectory data of the two vehicles constitute a lane-changing scenario. All lane-changing scenarios meeting the criteria are extracted, totaling 3600 scenario samples. If there are fewer than 100 lane-changing scenarios in an observed road segment, all data from the less than 50 frames before the lane change and the less than 50 frames after the lane change are extracted. This facilitates the analysis of the kinematic parameter distribution during the lane change process, as shown in Table 1 below. The filtered lane-changing scenarios are obtained as follows:

[0069] Table 1. Statistics of selected lane-change entry scenarios

[0070]

[0071] In this embodiment, data preprocessing is also performed on the selected lane-changing and merging scenarios and the corresponding driving scenario multidimensional datasets. Specifically, a three-step method of cleaning, alignment, and standardization can be used for data preprocessing. For example, timestamps are used for alignment to unify the spatial coordinates of the time-series data obtained by multiple sensors. At the same time, Gaussian filters or SG filters are used to remove noise and interference information from the data. For categorical data such as weather or road data, one-hot encoding is used to convert them into numerical data to unify the data type. Since the highD dataset used in this embodiment has already undergone high-precision extraction and post-processing using computer vision algorithms (such as U-Net detection network, RLS trajectory smoothing, etc.) during the acquisition process, the positioning error is usually less than 10 cm, and the data signal-to-noise ratio is extremely high. Therefore, no additional preprocessing scheme is required.

[0072] S3: Conduct a comprehensive cognitive risk assessment across both vertical and horizontal dimensions. Quantify the degree of danger of the scenario using the comprehensive cognitive risk score calculation formula, expressed as:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] in, This represents the overall perceived risk score. Headway represents the time required for a following vehicle to reach the current position of the preceding vehicle, and is used to measure driving efficiency; The longitudinally corrected collision time is equal to or less than the smallest root of the quadratic equation. The longitudinally corrected collision time is the root of the quadratic equation. It makes up for the shortcomings of the traditional longitudinal collision time index by taking into account the relative speed and acceleration of the front and rear vehicles, highlighting the urgency of the collision. That is, the smaller the longitudinally corrected collision time, the shorter the time before the collision occurs and the more dangerous the scenario. The lateral collision time is specifically defined as the ratio between the lateral distance between the two vehicles and their lateral relative velocity, and is used to measure the risk of lateral collision between the two vehicles during the cutting-in process. , These represent the longitudinal distance between the cold chain logistics vehicle and the target vehicle ahead, and the longitudinal speed of the cold chain logistics vehicle, respectively. , , , These are the length, longitudinal position, speed, and acceleration of the vehicle in front. , , These are the longitudinal position, speed, and acceleration of the following vehicle (i.e., the refrigerated logistics vehicle). The lateral distance between the two vehicles. The lateral relative speed between two vehicles is the relative speed at which the vehicle in front approaches the vehicle when it changes lanes. It is used to measure the time required for the two vehicles to approach each other laterally.

[0080] To determine the relative positions of the two vehicles in the coordinate system, the corresponding calculation formula is used, taking the direction of travel as the reference. In the top view coordinate system, the front vehicle is relatively above the rear vehicle. Therefore, the lateral distance is calculated using the following formula:

[0081] ;

[0082] If the vehicle in front is below the vehicle behind, the lateral distance is calculated using the following formula:

[0083] ;

[0084] in, , These are the horizontal positions of the upper left corner of the front vehicle. , These refer to the lateral position and width of the upper left corner of the rear vehicle;

[0085] S4: Construct a risk assessment system based on three-dimensional indicators and clustering mapping to achieve scenario risk classification;

[0086] In this embodiment, the three-dimensional risk indicators for each lane change / cut-in scenario are calculated at all times, and then a clustering algorithm is used based on... , , Three-dimensional metrics are used to cluster and classify all lane-changing entry scenarios. The number of clustering iterations is preferably set to N=100 to ensure convergence. Elbow method and profile coefficient can be used to evaluate the clustering effect.

[0087] The K-means algorithm can classify lane-change scenarios of different risks based on the data's own distribution characteristics, even when the threshold is unknown. It calculates the risk score of each lane-change scenario within a cluster and maps the risk level based on the clustering results. For example... Figure 2 As shown in Table 2, four clustering results were obtained: cluster 0, cluster 1, cluster 2, and cluster 3. These results represent the unsupervised clustering distribution statistics for lane-changing entry scenarios with different risk levels.

[0088] Table 2. Clustering Results Statistics

[0089]

[0090] This embodiment calculates the risk level based on the comprehensive cognitive risk score formula and combines clustering results with risk level mapping. Since the calculated lateral collision time (LTG) is positive for all scenarios, the lateral collision risk represented by LTG needs to be jointly judged in conjunction with MTTC / THW. If the LTG is positive for a lane-changing cut-in scenario, and the corresponding MTTC and THW are none / 0, then the risk of that scenario is dominated by LTG. It should be noted that if the LTG value in a scenario is significantly greater than the corresponding MTTC or THW value, it means that the lateral collision risk is low, and the risk of that scenario is mainly dominated by longitudinal risk indicators. The contribution of the item to the overall risk score RS is negligible; only the longitudinal risk item is retained for calculation.

[0091] ;

[0092] Table 3 below shows the statistical results of the average risk scores of the clusters:

[0093] Table 3. Statistics on the average risk scores of clusters

[0094]

[0095] Based on the definition and comparison of the average risk score of the clusters, cluster 2 < cluster 0 < cluster 1 < cluster 3. Cluster 0 and cluster 2 are classified into the same category. The higher the risk score, the greater the risk of the scenario. Thus, the lane-changing entry scenario data of the three risk categories can be obtained. Among them, the scenario belonging to cluster 3 is a high-risk lane-changing entry scenario, cluster 1 is a medium-risk scenario, and clusters 0 and 2 are low-risk scenarios. This realizes unsupervised clustering of lane-changing entry scenarios with different risks.

[0096] like Figure 3 As shown in Table 4, kinematic parameters such as longitudinal velocity, longitudinal acceleration, lateral velocity, and lateral acceleration were extracted for low, medium, and high-risk lane-changing scenarios. Statistical features were calculated using tools, including maximum, minimum, average, median, 5th percentile, 95th percentile, σ value, and the frequency of each value. The statistical results of the kinematic parameter features for lane-changing scenarios with different risks are shown in Table 4 below.

[0097] Table 4. Statistical table of kinematic parameter characteristics for lane change scenarios with different risks.

[0098]

[0099] According to the kinematic parameter feature distribution chart of different risk lane change scenarios, it can be seen that the performance of high-risk scenarios in terms of relevant kinematic parameter features is more aggressive than that of low-risk and medium-risk scenarios, which strongly verifies the effectiveness of the unsupervised risk clustering method for different lane change scenarios based on three-dimensional indicators proposed in this invention.

[0100] S5: Extract high-risk lane-changing entry scenario data obtained from clustering for high-risk entry behavior analysis. In this embodiment, 934 high-risk lane-changing entry scenarios were selected (accounting for 25.9%), including driving data of cold chain logistics vehicles and target vehicles, such as the speed, acceleration, and distance between the two vehicles; cold chain logistics vehicle operation data such as load and refrigeration power; and meteorological environmental data such as temperature, rainfall, and snowfall.

[0101] Furthermore, the data of high-risk lane-changing scenarios are normalized, including Z-Score standardization, maximum-minimum normalization, and nonlinear transformation (logarithmic transformation, exponential transformation) normalization. According to the specific data characteristics and model requirements, the corresponding normalization method is selected to eliminate the influence of units, thereby achieving the effect of cleaning data and removing noise.

[0102] S6: Construct a physical model of the cold chain logistics vehicle, considering the influence of load, refrigeration power, weather, etc., construct a two-dimensional physical model that couples dynamic characteristics with the environment, quantify the influence of cold chain characteristics on vehicle response, obtain the longitudinal dynamic equation of the cold chain logistics vehicle including the influence of load, refrigeration power, and precipitation, and obtain the acceleration and deceleration of the cold chain logistics vehicle.

[0103] Based on the dynamic characteristics model of cold chain logistics vehicles, the longitudinal dynamic equation is established as follows:

[0104] ;

[0105] in, This indicates the acceleration of the cold chain logistics vehicle. This indicates the longitudinal driving force of the vehicle (affected by precipitation). It is the acceleration due to gravity; This indicates the power consumption of the refrigeration system. Through a coefficient Transformed into longitudinal driving force The loss, i.e., additional resistance, The power loss coefficient of refrigeration capacity can be obtained by fitting the collected driving data of cold chain logistics vehicles. This indicates the current speed of the cold chain logistics vehicle. Road rolling resistance is positively correlated with load. Indicates the rolling resistance coefficient; The longitudinal wind resistance experienced by the vehicle;

[0106] Let the total mass of the cold chain logistics vehicle be... , is represented as:

[0107] ;

[0108] in, For the vehicle's own weight, For cargo mass, in the vehicle dynamics equations, mass affects the vehicle's inertia and forces, thus affecting acceleration and deceleration.

[0109] Considering weather conditions, primarily precipitation and longitudinal wind resistance, the longitudinal component of wind resistance on a vehicle is expressed as:

[0110] ;

[0111] in, For aerodynamic coefficients, This represents the current speed of the cold chain logistics vehicle. This refers to the vehicle's frontal area. air density, For wind speed, The angle between the wind direction and the vehicle's direction of travel;

[0112] To quantify the impact of precipitation on the acceleration of cold chain logistics vehicles, this embodiment first analyzes the relationship between precipitation and road surface adhesion coefficient. It's about precipitation. The function, With road surface adhesion coefficient Related to road surface adhesion coefficient It decreases as precipitation increases, therefore, It can be modeled as being related to the road surface adhesion coefficient. Related functions. Query experimental data on pavement adhesion coefficient under different precipitation levels, plot a scatter plot of precipitation versus adhesion coefficient, and verify the negative correlation between the two using the Pearson correlation coefficient. (Pavement adhesion coefficient) It is related to road surface water accumulation, which in turn is related to rainfall. Based on the physical law that the road surface adhesion coefficient decreases nonlinearly with increasing precipitation, an exponential decay model is constructed:

[0113] ;

[0114] in, It is the road adhesion coefficient of a dry road surface. It is the attenuation coefficient of the road surface adhesion coefficient with precipitation. . The coefficient of adhesion of dry asphalt pavement depends on the road surface material, tire characteristics, and the water film formation process. = 0.8, moderate rain (p=20mm / h) = 0.5. Substituting the above data into the nonlinear decay function, and combining multiple sets of measurement data under different precipitation intensities, the nonlinear least squares method is used for fitting to obtain... The typical value range is 0.017 to 0.025. This embodiment preferably uses... =0.02 is used as an example parameter. It can also be calibrated and adjusted according to specific test scenarios (such as specific road surfaces or specific tires);

[0115] The longitudinal dynamic equations of the cold chain logistics vehicle, i.e., the complete acceleration equations, are constructed as follows:

[0116] ;

[0117] The vehicle's acceleration is calculated based on the physical model of the cold chain logistics vehicle. The driving trajectory of the cold chain logistics vehicle and the target vehicle in the lane-changing scenario is generated based on the model of safe distance. The model assumes that when the target vehicle completes the lane change, the vehicle brakes to the same speed as the vehicle changing lanes with the maximum deceleration, and the two vehicles maintain the minimum safe distance.

[0118] Specifically, based on the distances traveled during uniform speed driving, variable deceleration driving, and uniform deceleration driving, and the length of the vehicle changing lanes, the minimum safe distance is calculated to determine the initial state of the cold chain logistics vehicle, i.e., its longitudinal position, lateral position, and initial speed. Starting from this initial state, the cold chain logistics vehicle decelerates at its maximum deceleration until its speed drops to the same speed as the target vehicle at the moment the lane change is completed. Figure 4 As shown, a cold chain logistics vehicle following trajectory is generated based on the target vehicle's driving trajectory. The longitudinal displacement of the cold chain logistics vehicle changes over time, while its lateral position remains unchanged.

[0119] S7: Construct a conditional vector c of the coupling characteristics of the fused environmental interference;

[0120] In this embodiment, six key parameters are selected: initial longitudinal velocity, initial lateral velocity, lateral displacement, lateral peak velocity, time frame corresponding to the lateral velocity peak, and road surface adhesion coefficient. A condition vector c is constructed, and the selected key parameters are standardized, such as by using the max-min normalization method to normalize each parameter to [0,1], so as to provide structured input for the generated model.

[0121] S8: Construct a conditional variational autoencoder that integrates physical constraints and a weighted multi-objective loss function;

[0122] In order to generate lane-changing trajectories that conform to both the distribution of real driving data and the basic motion laws of vehicles, this embodiment constructs a conditional variational autoencoder and embeds a weighted multi-objective loss function to systematically balance the physical feasibility, environmental adaptability and consistency with real driving behavior of the reconstructed trajectory sequence.

[0123] In this embodiment, the input is the trajectory sequence X and the condition vector c of a high-risk lane-change cutting scenario, and the output is the reconstructed trajectory sequence with the same dimension as the real trajectory. Simultaneously, the sequence of characteristic parameters such as lateral displacement and lateral velocity is decoupled;

[0124] Specifically, a feedforward neural network is used to build an encoder to process trajectory sequences from high-risk lane-change scenarios. With condition vector By using a stacked structure of fully connected layers and ReLU activation functions, deep fusion of temporal trajectory features and structured conditional features is achieved, outputting the posterior distribution parameters of the latent space, including the mean. and logarithmic variance To achieve the mapping between the original data and the latent space, preferably, the time step T=100 corresponds to a 4-second lane-changing process (frame rate 25Hz). The feature dimension D=3 represents the core parameters characterizing the lane-changing action: lateral lane-changing displacement, longitudinal velocity of the lane-changing vehicle, and lateral velocity of the lane-changing vehicle. The condition vector... These are the structured constraint parameters for controlling trajectory generation, followed by the introduction of standard normal distribution noise. ,pass Random sampling of latent vectors Perform model training and trajectory generation, construct a decoder network, and input latent vectors. and condition vector The concatenated data is fed into another feedforward neural network, and the decoded output reconstructs the trajectory sequence. The sequence is eventually decoupled into a kinematic feature sequence, namely, a sequence of lateral displacement, lateral velocity, longitudinal velocity, etc., which can be directly used for physical constraint verification.

[0125] This embodiment implements constraints through a weighted multi-objective loss function, and the total loss function... Represented as:

[0126] ;

[0127] Trajectory Reconstruction Loss The reconstructed trajectory sequence is calculated using mean square error. With the actual trajectory The difference lies in the fact that L2 loss is more sensitive to larger errors and can effectively drive the reconstructed trajectory sequence to approximate the real data in terms of overall shape, as expressed in:

[0128] ;

[0129] KL divergence loss Calculate the posterior distribution of the latent variables With standard normal prior The degree of similarity between them, and the distribution constraint loss of key motion features. To ensure that the reconstructed trajectory sequence is consistent with the real data in the statistical distribution of key motion features (such as extreme values ​​of lateral displacement and peak velocity), a constraint based on one-dimensional Wasserstein distance is introduced. This distance can more robustly measure distribution differences, is applicable to non-overlapping distributions, and reduces smoothness loss. Lateral velocity of trajectory sequence reconstructed using L2 loss constraints The variation amplitude of the first-order difference in the temporal sequence is intended to punish unreasonable sharp jitters or abrupt changes, improving the smoothness and comfort of the trajectory, and reducing physical consistency loss. The lateral displacement change of the trajectory sequence reconstructed using L2 loss constraints should be approximately equal to the discrete integral of the lateral velocity, expressed as:

[0130] ;

[0131] In this embodiment, by adjusting the weighting parameters , , , , It can precisely control the balance between multiple objectives such as diversity exploration, data fitting, motion smoothing, and physical rationality in the reconstructed trajectory sequence. In this embodiment, the preferred weighting parameters are: , , , , Under these parameters, the model achieves the optimal balance between trajectory reconstruction accuracy and generation diversity. The weighting coefficients for each loss term are set according to the following principles: Trajectory Reconstruction Loss The similarity between the generated trajectory and the real trajectory is directly determined, and it is the core optimization objective, therefore it is given the highest weight; KL divergence loss Regularization of the latent space is controlled by assigning a baseline weight of 1.0; distribution constraint loss. and smoothness loss Physical consistency loss As auxiliary constraints, the generation quality is improved from the dimensions of statistical distribution, trajectory smoothness and kinematic consistency, and their weights are relatively reduced to avoid dominating the main loss term. , , , , The value of is determined through ablation experiments. Specifically, a grid search method is used on the validation set, with the JS divergence and physical consistency error between the generated trajectory and the real trajectory as evaluation indicators. Combinations of coefficients within the range [0, 5.0] with intervals of 0.2 are tested, and the optimal weight parameter combination yields the smallest JS divergence and physical consistency error. In practical applications, the above coefficients can be adjusted appropriately according to different emphases on smoothness, distribution similarity, etc.; if the application scenario of the generated trajectory has high requirements for smoothness and comfort (such as scenarios sensitive to cargo stability), the coefficients can be appropriately increased. If a high coverage requirement for extreme operating conditions is needed, the coverage level can be appropriately reduced. This increases the diversity of latent variables. The weight configuration disclosed in this embodiment is a preferred scheme, which can effectively balance trajectory reconstruction accuracy and generation diversity.

[0132] In this embodiment, trajectory generation and optimization based on the fusion of data-driven and physical models deeply integrates the data-driven generation model with explicit physical prior knowledge to form a trajectory generation framework that combines learning and constraints, ensuring that the output trajectory has both the data characteristics of real driving behavior and conforms to vehicle dynamics constraints in different environments.

[0133] In this embodiment, the conditional variational autoencoder learns the feature distribution of high-risk lane-change trajectories through the encoder, and the decoder is based on the latent vector. Generate trajectories that conform to real-world distributions, achieving the goal of resembling real-world trajectories. This is based on rainfall data. Road surface adhesion coefficient As a regulating factor, according to The value is adjusted proportionally to the velocity variation component in the trajectory. The lower the adhesion coefficient, the stronger the suppression of velocity changes, thus automatically generating a more conservative and stable lane-changing trajectory adapted to slippery road surfaces. The reconstructed trajectory sequence is then reconstructed using normalized parameters. The inverse transformation is performed back to a trajectory with actual physical units, completing the output from the abstract feature space to the real physical world, and then a fifth-order polynomial fitting is performed:

[0134] ;

[0135] Among them, the independent variable This represents time step t, corresponding to a 4-second time series of the lane-changing process, with the dependent variable being... The characteristic parameters representing the trajectory describe the lateral position and can also be used for fitting longitudinal or lateral velocities. The fitting parameters of the polynomial are solved using the least squares method. , , , , , ,in, The overall trend of change in the dominant trajectory. To assist in adjusting the curvature changes of the trajectory. The main influence is the change in acceleration that affects the trajectory. Rate of change of velocity of associated trajectory The average velocity of the corresponding trajectory, The initial value corresponding to the trajectory;

[0136] This embodiment uses fifth-order polynomial fitting to effectively eliminate high-frequency noise that may remain in the generated model, ensure the continuity of the derivatives of the trajectory, meet the input requirements of the vehicle controller, and generate lane-changing trajectories that conform to both real driving behavior and vehicle dynamics. This effectively improves the diversity, smoothness, and physical feasibility of trajectory generation and enables deep integration of data distribution learning of the generated model and prior constraints of the physical model.

[0137] In this embodiment, the distribution characteristics of motion feature parameters of the reconstructed trajectory sequence generated by the conditional autoencoder are statistically analyzed. Specifically, the probability distribution of motion feature parameters is estimated based on the KDE nonparametric method, including the maximum value, minimum value, average value, 5th percentile, median, 95th percentile, σ value, and the frequency of each value. The distribution characteristics of kinematic feature parameters are compared with those of high-risk entry scenarios, as shown in Table 5 below. By comparing the two sets of feature parameter data, the effectiveness of the conditional generation model in generating scene trajectories is verified.

[0138] Table 5. Comparison of kinematic parameter characteristics of lane change scenarios generated by different models with the original data.

[0139]

[0140] The table continues as follows:

[0141]

[0142] S9: Based on the difference in motion feature parameter distribution between the reconstructed trajectory sequence generated by the conditional variational autoencoder with multi-index joint evaluation and the real high-risk lane change and cutting scenario, an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles is constructed, which also provides a reference for the optimization of the scenario generation model structure.

[0143] Specifically, this embodiment evaluates the reconstructed trajectory sequence from multiple aspects, including distribution similarity, edge scene generation capability, and the impact of model input and loss function design on the quality of the reconstructed trajectory sequence.

[0144] In this embodiment, a distribution similarity index is used to measure the difference between the feature distribution of the reconstructed trajectory sequence generated by the conditional variational autoencoder and the true trajectory distribution. The distribution similarity index includes JS divergence and Wasserstein distance.

[0145] As shown in Table 6 below, JS divergence is used to calculate the three kinematic feature parameters of longitudinal velocity, lateral displacement, and lateral velocity of the highD dataset and the conditional variational autoencoder to evaluate the distribution similarity between the reconstructed trajectory sequence and the real trajectory.

[0146] Table 6. JS divergence measures the difference in parameter distribution between the true trajectory and the reconstructed trajectory sequence.

[0147]

[0148] The smaller the JS divergence value, the closer the reconstructed trajectory sequence is to the real trajectory distribution, and the better the model performance. The JS divergence quantitative evaluation results show that both types of conditional generation models exhibit good generation performance, and the generated trajectory data has a high similarity to the real data distribution. By comparing the JS divergence of lane-change trajectory data generated by variational autoencoder models based on different conditional dimensions, increasing the conditional dimension can significantly improve the generation quality of lateral displacement and lateral velocity, while the generation of longitudinal velocity is mainly affected by the initial longitudinal velocity and is not sensitive to additional conditions. Among them, the distribution fitting performance of the environmental conditional generation model considering the influence of precipitation is slightly lower than that of the distribution fitting performance of the generation model without environmental conditions. The reason for the above difference is that the highD dataset is collected in a dry environment, and the precipitation factor introduced in the environmental conditional generation model will affect the vehicle speed, resulting in the generated longitudinal velocity and lateral velocity distribution being slightly lower than the real data. This phenomenon precisely verifies the necessity of incorporating environmental factors into the scene generation model, which can generate driving trajectories that are more in line with the real physical world.

[0149] By adding key conditional features, the model can more accurately capture the distribution characteristics of real trajectories. The consistency between the reconstructed trajectory sequence and the distribution of real trajectories is significantly improved, which verifies the positive effect of conditional constraints on distribution matching. It solves the problem of insufficient generation of benchmark generation models in specific scenarios and insufficient physical rationality of scenarios, and further improves the diversity and realism of generated scenarios to meet the high-quality generation requirements of edge scenarios in autonomous driving testing and verification.

[0150] As shown in Table 7 below, the quantile error index is used to calculate and compare the differences in the distribution of feature parameters (such as acceleration, velocity, etc.) between the real high-risk entry scene and the reconstructed trajectory sequence at the 25th, 50th, and 75th quantiles. This reflects the edge scene generation capability and realism. Quantile error is an important indicator for evaluating the performance of the generation model. It can reflect the model's fitting accuracy at different positions (low, medium, and high quantiles) of the data distribution, making up for the limitations of using only the mean or standard deviation. The smaller the error value, the better the consistency between the generated data and the real data at that quantile.

[0151] Table 7. Schematic diagram of the trajectory data generation capability of the model generated under the evaluation conditions of quantile error index.

[0152]

[0153] Where y represents the lateral lane change displacement, vx represents the longitudinal velocity of the lane change vehicle, and vy represents the lateral velocity of the lane change vehicle;

[0154] From the quantile statistics, the environmental condition generation model performed better overall. The error of the environmental condition generation model was significantly lower than that of the model without environmental conditions at the quantiles of the three evaluation features. However, the environmental condition generation model had a local advantage. The error of the low and middle quantiles of the lateral velocity both exceeded 5%, which was mainly affected by the change in the road surface adhesion coefficient due to precipitation. When the road surface adhesion coefficient decreased, it would affect the magnitude of the longitudinal and lateral velocities.

[0155] In this embodiment, qualitative analysis indicators are used to qualitatively analyze the differences in trajectory generation quality caused by the number of conditional parameters and the design of the loss function in the conditional generation model. Through joint evaluation of multiple quantitative and qualitative indicators, the quality of the trajectory generated by the model can be effectively assessed, providing a theoretical basis for optimizing trajectory generation models in cold chain scenarios and further improving the quality and practicality of these models. By analyzing and verifying the rationality and realism of the scenarios using indicators from different dimensions, an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles is ultimately constructed.

[0156] As shown in Table 8 below, qualitative analysis of the model clarifies the differential impact of conditional parameters and loss function design on the quality of the reconstructed trajectory sequence:

[0157] Table 8. Analysis of the Influence Mechanism between Conditional Generation Model Parameter Conditions and Loss Function Design

[0158]

[0159] Where vx0 represents the initial longitudinal velocity and vy0 represents the initial lateral velocity. This represents the time frame corresponding to the peak lateral velocity. Indicates the maximum lateral velocity; This represents the lateral displacement during lane changing, where μ represents the road surface adhesion coefficient. Indicates the trajectory reconstruction loss. This represents the KL divergence loss. The distribution-constrained loss represents the key motion features. Indicates smoothness loss. This represents the loss of physical consistency.

[0160] First, the Conditional Model 2 series, which includes peak lateral velocity and lateral displacement, can more accurately constrain the lateral motion characteristics of the reconstructed trajectory sequence compared to the Conditional Model 1 series, improving the geometric accuracy and physical plausibility of the trajectory. Second, the environmental conditional model incorporating precipitation, in terms of distribution constraint loss... Under its influence, diverse trajectories with environmental adaptability can be generated, but this needs to be combined with physical consistency loss. Ensure the dynamics of the trajectory are realistic.

[0161] Qualitative analysis of the parameters of the conditional generation model reveals that a single-dimensional evaluation criterion cannot comprehensively and accurately reflect the quality of the reconstructed trajectory sequence. This analysis exposes the impact mechanism of conditional dimensions and loss function design on the quality of the reconstructed trajectory sequence. By introducing explicit conditional vectors and multiple loss function constraints, the conditional generation model significantly improves the controllability, realism, and scene adaptability of the reconstructed trajectory sequence, providing a more reliable technical solution for trajectory generation in cold chain scenarios. Further optimization of the conditional vector design and loss function weight allocation can achieve more accurate scene trajectory generation.

[0162] Example 2

[0163] This embodiment provides a scenario generation system for testing and evaluating autonomous cold chain logistics vehicles, which is used to execute the scenario generation method for testing and evaluating autonomous cold chain logistics vehicles in Embodiment 1. The system includes: a data acquisition module, a scenario filtering module, a risk assessment module, a trajectory extraction module, a driving trajectory generation module, a trajectory sequence reconstruction module, and a scenario library construction module.

[0164] In this embodiment, the data acquisition module is used to collect a multidimensional dataset of the driving scenario of the cold chain logistics vehicle;

[0165] In this embodiment, the scene filtering module is used to filter lane change entry scenes based on preset scene filtering rules;

[0166] In this embodiment, the risk assessment module is used to calculate the comprehensive cognitive risk score of each lane change scenario based on the headway, longitudinal correction collision time, and lateral collision time, and to cluster and classify all lane change scenarios to obtain the risk level of the lane change scenario.

[0167] In this embodiment, the trajectory extraction module is used to extract trajectory sequences of high-risk lane-changing cutting scenarios;

[0168] In this embodiment, the driving trajectory generation module is used to generate the driving trajectory of the cold chain logistics vehicle and the target vehicle in a lane-changing scenario, specifically including:

[0169] A physical model of the cold chain logistics vehicle is constructed based on its dynamic characteristics, and the longitudinal dynamic equation of the cold chain logistics vehicle is obtained. The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation, and the driving trajectory of the cold chain logistics vehicle in the lane-changing scenario with the target vehicle is generated.

[0170] In this embodiment, the trajectory sequence reconstruction module is used to input the conditional vector and the trajectory sequence of the high-risk lane change scene into the conditional variational autoencoder to obtain the reconstructed trajectory sequence and decouple to obtain motion feature parameters.

[0171] In this embodiment, the scenario library construction module is used to evaluate the difference in motion feature parameter distribution between the reconstructed trajectory sequence and the real high-risk lane change and cutting scenario, and to construct an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles.

[0172] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A scenario generation method for testing and evaluating autonomous cold chain logistics vehicles, characterized in that, Includes the following steps: Collect a multidimensional dataset of driving scenarios for cold chain logistics vehicles, and filter lane-changing entry scenarios based on preset scenario filtering rules; The comprehensive cognitive risk score for each lane change scenario is calculated based on the headway, longitudinal correction collision time, and lateral collision time. All lane change scenarios are clustered and classified to obtain the risk level of each lane change scenario. The trajectory sequence of high-risk lane change scenarios is extracted. Based on the dynamic characteristics of cold chain logistics vehicles, a physical model of the cold chain logistics vehicle is constructed, and the longitudinal dynamic equation of the cold chain logistics vehicle is obtained, which is expressed as: ； in, This indicates the acceleration of the cold chain logistics vehicle. This indicates the longitudinal driving force of the vehicle. It is the acceleration due to gravity. The exponential decay model represents the road surface adhesion coefficient. Indicates precipitation. This indicates the power consumption of the refrigeration system. Through loss coefficient Transformed into longitudinal driving force The loss, This is the power loss coefficient of cooling capacity. This indicates the current speed of the cold chain logistics vehicle. Road rolling resistance is positively correlated with load. Indicates the rolling resistance coefficient. The longitudinal wind resistance experienced by the vehicle; The total mass of the cold chain logistics vehicle is , is represented as: ； in, For the vehicle's own weight, For the quality of goods; The longitudinal wind resistance experienced by a vehicle is expressed as: ； in, For aerodynamic coefficients, This refers to the vehicle's frontal area. air density, For wind speed, The angle between the wind direction and the vehicle's direction of travel; The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation, and the driving trajectory of the cold chain logistics vehicle and the target vehicle in the lane-changing scenario is generated. Construct a conditional vector, input the conditional vector and the trajectory sequence of the high-risk lane change scenario into a conditional variational autoencoder to obtain a reconstructed trajectory sequence, and decouple to obtain motion feature parameters; The differences in motion characteristic parameter distribution between the reconstructed trajectory sequence and real high-risk lane change and cutting scenarios were evaluated, and an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles was constructed.

2. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 1, characterized in that, The multidimensional dataset for cold chain logistics vehicle driving scenarios includes cold chain logistics vehicle operation data, meteorological environment data, and driving behavior data.

3. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 1, characterized in that, The preset scenario selection rules are: the cold chain logistics vehicle is the main vehicle, the target lane of the vehicle in front to change lanes is the lane where the cold chain logistics vehicle is located, and after changing lanes, it is in front of the cold chain logistics vehicle, and there are no other vehicles between the two vehicles.

4. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 1, characterized in that, The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation and expressed as: ； in, This indicates the acceleration of the cold chain logistics vehicle. This represents the road adhesion coefficient of a dry road surface. It is the attenuation coefficient of the road surface adhesion coefficient with precipitation. Indicates precipitation. It is the acceleration due to gravity. This is the power loss coefficient of cooling capacity. For the power consumed by the cooling system, This indicates the current speed of the cold chain logistics vehicle. For the total mass of the cold chain logistics vehicle, Indicates the rolling resistance coefficient. This refers to the longitudinal wind resistance experienced by the vehicle.

5. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 1, characterized in that, Generate the driving trajectory of the cold chain logistics vehicle and the target vehicle in the lane-changing scenario, specifically including: The minimum safe distance is calculated based on the distance traveled during constant speed driving, variable deceleration driving, and constant deceleration driving, as well as the length of the vehicle changing lanes. The longitudinal position, lateral position, and initial speed of the cold chain logistics vehicle in the initial state are then determined. The cold chain logistics vehicle starts from the initial state and decelerates at maximum speed until its speed drops to the same speed as the target vehicle at the moment of lane change. The cold chain logistics vehicle then generates a following trajectory based on the target vehicle's trajectory.

6. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 1, characterized in that, Constructing conditional vectors specifically includes: A condition vector is constructed by selecting the initial longitudinal velocity, initial lateral velocity, lateral displacement, peak lateral velocity, the time frame corresponding to the peak lateral velocity, and the road adhesion coefficient.

7. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 1, characterized in that, The conditional vector and the trajectory sequence of the high-risk lane-change scenario are input into a conditional variational autoencoder to obtain a reconstructed trajectory sequence, specifically including: The conditional vector and the trajectory sequence of the high-risk lane-changing cutting scenario are encoded by the posterior distribution parameters of the latent space, and the latent vector is randomly sampled based on the posterior distribution parameters and noise. The latent vector and the conditional vector are concatenated and input into the decoder, and the decoded output is a reconstructed trajectory sequence.

8. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 7, characterized in that, In the step of reconstructing the trajectory sequence, a total loss function is constructed based on trajectory reconstruction loss, KL divergence loss, distribution constraint loss, smoothness loss, and physical consistency loss; The trajectory reconstruction loss is calculated using the mean squared error to determine the difference between the reconstructed trajectory sequence and the true trajectory. KL divergence loss calculates the similarity between the posterior distribution of a latent variable and the standard normal prior. The distribution constraint loss, based on one-dimensional Wasserstein distance constraints, reconstructs the motion feature statistical distribution of the trajectory sequence, which is consistent with the true trajectory. The smoothness loss uses L2 loss constraints to reconstruct the magnitude of the lateral velocity variation of the trajectory sequence in the first difference over time. The change in lateral displacement of the reconstructed trajectory sequence using L2 loss constraint is approximately equal to the discrete integral of the lateral velocity.

9. The scenario generation method for testing and evaluating autonomous cold chain logistics vehicles according to claim 7, characterized in that, The process of reconstructing the trajectory sequence also includes trajectory adjustment and fitting steps, specifically including: The velocity variation component of the reconstructed trajectory sequence is adjusted proportionally based on the road surface adhesion coefficient. Polynomial fitting is performed on the reconstructed trajectory sequence.

10. A scenario generation system for testing and evaluating autonomous cold chain logistics vehicles, characterized in that, The scenario generation method for performing the test evaluation of autonomous driving cold chain logistics vehicles according to any one of claims 1-9 includes: a data acquisition module, a scenario screening module, a risk assessment module, a trajectory extraction module, a driving trajectory generation module, a trajectory sequence reconstruction module, and a scenario library construction module. The data acquisition module is used to collect multidimensional datasets of cold chain logistics vehicle driving scenarios; The scene filtering module is used to filter lane-change entry scenes based on preset scene filtering rules; The risk assessment module is used to calculate the comprehensive cognitive risk score of each lane change scenario based on the headway, longitudinal correction collision time, and lateral collision time, and to cluster and classify all lane change scenarios to obtain the risk level of the lane change scenario. The trajectory extraction module is used to extract trajectory sequences for high-risk lane-changing and cutting-in scenarios; The driving trajectory generation module is used to generate the driving trajectory of the cold chain logistics vehicle in the lane-changing scenario with the target vehicle, specifically including: A physical model of the cold chain logistics vehicle is constructed based on its dynamic characteristics, and the longitudinal dynamic equation of the cold chain logistics vehicle is obtained. The acceleration of the cold chain logistics vehicle is calculated based on the longitudinal dynamic equation, and the driving trajectory of the cold chain logistics vehicle in the lane-changing scenario with the target vehicle is generated. The trajectory sequence reconstruction module is used to input the conditional vector and the trajectory sequence of the high-risk lane change scene into the conditional variational autoencoder to obtain the reconstructed trajectory sequence and decouple to obtain motion feature parameters. The scenario library construction module is used to evaluate the differences in motion feature parameter distribution between the reconstructed trajectory sequence and the real high-risk lane change and cutting scenario, and to construct an autonomous driving test scenario library based on the characteristics of cold chain logistics vehicles.

Images