Vehicle test strategy generation method and device, storage medium and equipment

Abstract

The invention provides a vehicle test strategy generation method and device, a storage medium and equipment, and the method comprises the steps: obtaining feature scene data which is not covered by a standard scene library in a target region, including a non-standard road mark, a traffic participant behavior mode and an endemic weather combination, training a target large model based on the feature scene data, and obtaining a target test strategy; according to the method, a target large model learns scene feature vectors extracted from input data in the training process, then a test strategy is generated according to the scene feature vectors, and after training is completed, a target test strategy corresponding to a test scene is automatically generated through the model. Thus, by collecting the characteristic scene data and constructing the scene library covering the regional characteristics, the trained target large model can automatically generate the test strategy covering the regional characteristics, and the reliability of the test result is improved.

Description

Technical Field

[0001] This application relates to the field of vehicle testing technology, and more specifically, to a method, apparatus, storage medium, and device for generating vehicle testing strategies. Background Technology

[0002] As autonomous driving technology advances towards L4 / L5 level commercialization, the comprehensiveness and scenario coverage of testing strategies have become core aspects of ensuring system safety. Traditional test strategy generation mainly relies on standard scenario libraries. However, standard scenario libraries are based on European and American road environments and lack sufficient coverage of regionally specific scenarios, which can easily lead to test results that fail to reflect real-world, complex environments. Summary of the Invention

[0003] The purpose of this application is to provide a vehicle testing strategy generation method, apparatus, storage medium and device, which aims to solve the problem that the vehicle testing strategy generation methods in the related technology do not cover regional characteristic scenarios enough, resulting in the test results failing to reflect the real complex environment.

[0004] In a first aspect, this application provides a vehicle testing strategy generation method, comprising: acquiring characteristic scene data in a target area that is not covered by a standard scene library; the characteristic scene data including non-standard road markings, traffic participant behavior patterns, and local weather combinations; training a target large model based on the characteristic scene data to obtain a trained target large model; during the training process, the target large model learns to extract scene feature vectors from input data, and then generates a testing strategy based on the scene feature vectors; and generating a target testing strategy corresponding to the test scene based on the trained target large model.

[0005] In the above implementation process, distinctive scene data not covered by the standard scene library in the target area is acquired, including non-standard road markings, traffic participant behavior patterns, and local weather combinations. Then, a target large-scale model is trained based on this distinctive scene data. During training, the target large-scale model learns to extract scene feature vectors from the input data and generates test strategies based on these feature vectors. After training, the model automatically generates target test strategies corresponding to the test scenarios. In this way, by collecting distinctive scene data and constructing a scene library covering regional characteristics, the trained target large-scale model can automatically generate test strategies covering regional characteristics, improving the reliability of test results.

[0006] Furthermore, in some examples, the non-standard road markings include hand-drawn lane lines and temporary obstacles; the acquisition of characteristic scene data in the target area not covered by the standard scene library includes: collecting image data and point cloud data containing non-standard road markings using a test vehicle traveling in the target area; extracting the hand-drawn lane lines from the image data using an adaptive threshold Canny algorithm, and labeling the image data based on the hand-drawn lane lines; identifying the temporary obstacles from the point cloud data using a DBSCAN clustering algorithm, and labeling the point cloud data based on the temporary obstacles.

[0007] In the above implementation process, a specific method is provided to obtain non-standard road markings in the target area that are not covered by the standard scene library. This avoids the risk that the model will ignore the risk of construction area due to not having encountered non-standard road markings, and thus generate invalid test strategies.

[0008] Furthermore, in some examples, the acquisition of characteristic scene data in the target area that is not covered by the standard scene library also includes: collecting non-standard traffic trajectory data by using test vehicles traveling in the target area; simulating pedestrian and vehicle movement using an extended social force model based on the non-standard traffic trajectory data to obtain the traffic participant behavior pattern; and recording the traffic participant behavior pattern in text data.

[0009] In the above implementation process, while the test vehicle is driving in the target area, it collects non-standard traffic trajectory data such as disorderly lane changes by motorcycles and pedestrians crossing the road. Then, an extended social force model is used to simulate pedestrian / vehicle movements to obtain the behavioral patterns of traffic participants, which are recorded in text data to provide structured input for subsequent model training. In this way, the model's ability to understand, predict, and make decisions in complex dynamic scenarios is improved.

[0010] Furthermore, in some examples, acquiring characteristic scene data in the target area that is not covered by the standard scene library also includes: acquiring real-time meteorological data of the target area from a weather station; the real-time meteorological data includes rainfall, visibility, and wind speed; acquiring environmental data from the onboard sensors of a test vehicle located in the target area; combining the real-time meteorological data and the environmental data to obtain the local weather combination; and recording the local weather combination in the text data.

[0011] In the above implementation process, real-time data such as rainfall, visibility, and wind speed are obtained from weather stations, and environmental data such as road surface humidity and temperature are obtained from vehicle-mounted sensors. This forms local weather combinations, captures the interaction effects between weather and environment, and after analyzing the local weather combinations, they are recorded in text data to provide structured input for subsequent model training. In this way, the testing strategy generated by the model is more adapted to local weather and environment.

[0012] Furthermore, in some examples, the target large model includes a combined scene generation sub-model and a test strategy generation sub-model; the combined scene generation sub-model includes a multimodal encoder, a cross-modal attention layer, and a dynamic scene generator; the multimodal encoder is used to process the input image data, point cloud data, and text data respectively, and output corresponding feature vectors; the cross-modal attention layer is used to fuse feature vectors from different modalities to generate fused features; the dynamic scene generator is used to generate a combined scene based on the fused features; the test strategy generation sub-model is used to generate a test strategy based on the scene feature vectors corresponding to the combined scene; the test strategy includes the following parameters: sensor calibration frequency, path planning optimization step size, and environmental adaptability test threshold.

[0013] In the above implementation process, the target large model includes two sub-models. One sub-model comprises three modules: a multimodal encoder, a cross-modal attention layer, and a dynamic scene generator. The multimodal encoder processes image, point cloud, and text data respectively, outputting feature vectors for each modality. The cross-modal attention layer fuses features from different modalities, and the dynamic scene generator generates combined scenes based on the fused features. The other sub-model automatically generates test strategies covering regional characteristics based on the scene feature vectors output by the first sub-model. This overcomes the single-modal limitations of traditional models and effectively generates test strategies that conform to real-world, complex environments.

[0014] Furthermore, in some examples, the featured scene data also includes dynamic scene data triggered by cultural activities; the dynamic scene data includes traffic flow changes; the acquisition of featured scene data in the target area that is not covered by the standard scene library includes: acquiring the time and location of cultural activities in the target area; acquiring road closure information during the activity based on the time and location; and using the road closure information to predict traffic flow changes during the activity based on the ARIMA model.

[0015] In the above implementation process, the time and location of cultural activities such as religious festivals and markets are obtained to acquire road closure information during these activities. Then, the ARIMA model is used to predict traffic flow changes during these activities. This allows the trained target model to accurately simulate the dynamic scenarios triggered by cultural activities in the target area, improving model performance.

[0016] Furthermore, in some examples, the featured scene data also includes featured infrastructure; the featured infrastructure includes narrow dirt roads and temporary construction areas; the acquisition of featured scene data in the target area not covered by the standard scene library includes: acquiring basic road data of the target area based on a high-precision map of the target area; extracting images containing narrow dirt roads and temporary construction areas in the target area from UAV aerial photography data based on the basic road data; and performing cubic B-spline curve fitting on the images to obtain the boundaries of the narrow dirt roads and temporary construction areas.

[0017] In the above implementation process, based on the basic road data of the target area, images containing narrow dirt roads and temporary construction areas are extracted from UAV aerial photography data. Then, cubic B-spline curve fitting is used to generate a smooth and continuous boundary model. In this way, the gap in the standardized database regarding specialized infrastructure is filled, further improving the model performance.

[0018] Furthermore, in some examples, the featured scene data also includes target traffic regulation clauses; the acquisition of featured scene data in the target area that is not covered by the standard scene library includes: using a pre-trained BERT model to extract the target traffic regulation clauses from text data containing local traffic regulations of the target area.

[0019] In the aforementioned implementation process, specific regulations are obtained from government documents and legal texts of the target region, and detailed implementation rules are obtained from traffic management departments. A pre-trained BERT model is used to extract key clauses of the regulations and record them into the text data to train the target large model. In this way, the final generated testing strategy can adapt to regional regulatory differences.

[0020] Furthermore, in some examples, it also includes: verifying the target testing strategy in a simulation environment and obtaining simulation verification metrics; the simulation verification metrics include path planning success rate and emergency braking response time.

[0021] In the above implementation process, after the strategy is generated, it can be verified in a simulation environment. The verification indicators include path planning success rate and emergency braking response time. In this way, the effect of the strategy can be accurately evaluated through simulation verification, which can promote strategy optimization.

[0022] Furthermore, in some examples, the method also includes: verifying the target testing strategy in a real vehicle environment to obtain real vehicle verification results; and iteratively optimizing the dynamic scene generator by feeding back the simulation verification metrics and the real vehicle verification results to the target large model.

[0023] In the above implementation process, after completing the real-vehicle verification, the real-vehicle verification results, together with the preceding simulation verification indicators, are fed back to the model to iteratively optimize the dynamic scene generator. In this way, by optimizing model parameters through simulation and real-vehicle verification feedback, the entire process from data acquisition to strategy iteration is automated, effectively improving the reliability of the model.

[0024] Secondly, this application provides a vehicle testing strategy generation device, comprising: an acquisition module for acquiring characteristic scene data in a target area that is not covered by a standard scene library; the characteristic scene data includes non-standard road markings, traffic participant behavior patterns, and local weather combinations; a training module for training a target large model based on the characteristic scene data to obtain a trained target large model; during the training process, the target large model learns to extract scene feature vectors from input data and then generates a testing strategy based on the scene feature vectors; and a generation module for generating a target testing strategy corresponding to a test scene based on the trained target large model.

[0025] Thirdly, this application provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method described in any of the first aspects.

[0026] Fourthly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described in any of the first aspects.

[0027] Fifthly, this application provides a computer program product that, when run on a computer, causes the computer to perform the method described in any of the first aspects.

[0028] Other features and advantages disclosed in this application will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the above-described technology disclosed in this application.

[0029] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart illustrating a vehicle testing strategy generation method provided in this application embodiment; Figure 2 A schematic diagram illustrating the workflow of an intelligent test strategy automatic generation scheme based on a large model, provided in an embodiment of this application; Figure 3 A block diagram of a vehicle testing strategy generation device provided in an embodiment of this application; Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0033] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0034] The comprehensiveness and scenario coverage of testing strategies are key elements in ensuring the safety of autonomous driving systems. Traditional test strategy generation mainly relies on standard scenario libraries, such as Waymo Open Dataset and CARLA built-in scenarios. These standard scenario libraries provide the data foundation for testing strategies and ensure the reliability of testing. However, standard scenario libraries mostly focus on standardized road environments in Europe and America, lacking modeling for regionally specific scenarios, which can easily lead to test results that fail to reflect real-world, complex environments.

[0035] To address the aforementioned issues, this application provides a vehicle testing strategy generation scheme. It acquires characteristic scene data from a target area not covered by a standard scene library, including non-standard road markings, traffic participant behavior patterns, and local weather combinations. Then, based on this characteristic scene data, a target large-scale model is trained. This model learns to extract scene feature vectors from the input data and generates testing strategies based on these feature vectors. After training, the model automatically generates target testing strategies corresponding to the test scenarios. In this way, by collecting characteristic scene data and constructing a scene library covering regional characteristics, the trained target large-scale model can generate dynamic scenarios that better reflect real-world logic, thereby generating more effective testing strategies and improving the reliability of test results.

[0036] The embodiments of this application will be described below: like Figure 1 As shown, Figure 1 This is a flowchart illustrating a vehicle testing strategy generation method provided in an embodiment of this application. The method includes: Step 101: Obtain characteristic scene data in the target area that is not covered by the standard scene library; the characteristic scene data includes non-standard road markings, traffic participant behavior patterns, and local weather combinations; The target area mentioned in this step can be a road environment that differs from the standardized road environment in Europe and America, such as rural roads in Southeast Asia or urban roads in India. This embodiment collects characteristic scene data such as non-standard road markings, traffic participant behavior patterns, and local weather combinations within the target area to construct a scene library covering regional differences, laying a solid data foundation for subsequent training of large models capable of generating dynamically combined scenes.

[0037] In some embodiments, the non-standard road markings mentioned in this step may include hand-drawn lane lines and temporary obstacles; then this step may include: acquiring image data and point cloud data containing non-standard road markings using a test vehicle traveling in the target area; extracting the hand-drawn lane lines from the image data using an adaptive threshold Canny algorithm, and labeling the image data based on the hand-drawn lane lines; identifying the temporary obstacles from the point cloud data using a DBSCAN clustering algorithm, and labeling the point cloud data based on the temporary obstacles. That is, a camera and LiDAR can be installed on the test vehicle. When the test vehicle travels in the target area, it acquires image data and point cloud data containing hand-drawn lane lines and temporary obstacles; then, for the image data, the hand-drawn lane lines are extracted using an adaptive threshold Canny algorithm, and polygonal regions are labeled using these; for the point cloud data, the temporary obstacles, such as bamboo scaffolding, are identified using a DBSCAN clustering algorithm, and cubebox labels are performed on the point cloud data, providing structured input for subsequent model training. This avoids the risk that the model might overlook construction area risks due to a lack of experience with non-standard road markings, thus generating invalid testing strategies.

[0038] In some embodiments, this step may further include: collecting non-standard traffic trajectory data using a test vehicle traveling in the target area; simulating pedestrian and vehicle movement using an extended social force model based on the non-standard traffic trajectory data to obtain the traffic participant behavior pattern; and recording the traffic participant behavior pattern in text data. That is, millimeter-wave radar can be installed around the test vehicle. When the test vehicle travels in the target area, it collects non-standard traffic trajectory data. Here, non-standard traffic trajectory data refers to trajectory data that deviates from conventional traffic behavior patterns, such as disorderly lane changes by motorcycles or pedestrians crossing the road. Then, based on the collected data, an extended social force model is used to simulate pedestrian / vehicle movement. The social force model is a classic model describing pedestrian / vehicle movement, and its core assumption is that individual movement is driven by both "attraction" (such as a target point) and "repulsion" (such as obstacles or other individuals). Considering the insufficient adaptability of traditional social force models to disordered behavior, an extended social force model is introduced to construct a dynamic pedestrian / vehicle motion simulation framework. Optionally, this extended social force model can describe individual motion through goal-oriented forces, interaction forces, and environmental forces. Here, goal-oriented forces can refer to the "attraction" that drives individuals to move towards a target point, reflecting the purposefulness of behavior. Interaction forces can refer to the "repulsion" or "following force" generated by spatial competition between individuals, reflecting social behavioral rules. Environmental forces can refer to the constraints of the physical environment, such as road boundaries and obstacles, on individual motion. The motion behavior simulated by this extended social force model is the behavior pattern of traffic participants, which is recorded in text data to provide structured input for subsequent model training. In this way, the model's ability to understand, predict, and make decisions in complex dynamic scenarios is improved.

[0039] In some embodiments, this step may further include: acquiring real-time meteorological data of the target area from a meteorological station; the real-time meteorological data includes rainfall, visibility, and wind speed; acquiring environmental data from onboard sensors of a test vehicle located in the target area; combining the real-time meteorological data and the environmental data to obtain the local weather combination; and recording the local weather combination in text data. That is, real-time data such as rainfall, visibility, and wind speed can be acquired from a meteorological station, and environmental data such as road surface humidity and temperature can be acquired from onboard sensors to form a local weather combination, capturing the interaction between weather and environment. For example, the rainfall-road surface humidity combination can quantify the water absorption rate of the road surface under different rainfall intensities, and the wind speed-visibility combination can analyze the diffusion effect of strong winds on dust and fog, correcting visibility predictions. Similarly, after analyzing and obtaining the local weather combination, it is recorded in text data to provide structured input for subsequent model training. This makes the model-generated testing strategy more adaptable to local weather and environment.

[0040] Step 102: Train the target large model based on the characteristic scene data to obtain the trained target large model; during the training process, the target large model learns to extract scene feature vectors from the input data, and then generates a test strategy based on the scene feature vectors. The target large model mentioned in this step is the large model used to automatically generate test strategies. After completing the collection and storage of characteristic scene data, the model is trained based on the collected characteristic scene data, enabling the model to learn to extract scene feature vectors from the input data, and then match appropriate test strategies according to the scene feature vectors.

[0041] In some embodiments, the target large model mentioned in this step may include a combined scene generation sub-model and a test strategy generation sub-model; the combined scene generation sub-model includes a multimodal encoder, a cross-modal attention layer, and a dynamic scene generator; the multimodal encoder is used to process the input image data, point cloud data, and text data respectively, and output corresponding feature vectors; the cross-modal attention layer is used to fuse feature vectors from different modalities to generate fused features; the dynamic scene generator is used to generate a combined scene based on the fused features; and the test strategy generation sub-model is used to generate a test strategy based on the scene feature vectors corresponding to the combined scene.

[0042] In other words, the target large model can include two sub-models. One sub-model can adopt the Transformer Encoder-Decoder structure, including three modules: a multimodal encoder, a cross-modal attention layer, and a dynamic scene generator. The multimodal encoder processes image, point cloud, and text data respectively, outputting feature vectors for each modality. The cross-modal attention layer fuses features from different modalities, and the dynamic scene generator generates combined scenes based on the fused features, such as motorcycles and pedestrians mixing in a rainstorm. The other sub-model automatically generates a test strategy covering regional characteristics based on the scene feature vector output by the first sub-model. This scene feature vector can include the following dimensions: non-standard road markings, traffic participant behavior patterns, and local weather combinations. The test strategy can be represented as a set of strategy parameters, such as: sensor calibration frequency, path planning optimization step size, and environmental adaptability test threshold. This overcomes the single-modal limitations of traditional models and effectively generates test strategies that conform to real-world complex environments.

[0043] Step 103: Generate a target testing strategy corresponding to the test scenario based on the trained target model.

[0044] This step refers to the process of inputting test scenario data into the trained target model after training the target model, in order to generate the corresponding target testing strategy. The test scenario can be a simulation test scenario or a real-vehicle test scenario.

[0045] To enhance the performance of the model, this application also provides the following improvements: In some embodiments, the featured scene data mentioned in step 101 may further include dynamic scene data triggered by cultural activities; the dynamic scene data includes traffic flow changes; then step 101 may further include: obtaining the time and location of cultural activities in the target area; obtaining road closure information during the activity based on the time and location; and using the road closure information to predict traffic flow changes during the activity based on an ARIMA model. That is, the time and location of cultural activities such as religious festivals and markets can be obtained from the government's public calendar for the target area, and road closure information during the activity can be obtained from traffic management departments. Then, using the obtained information, traffic flow changes during the activity are predicted based on an ARIMA (Autoregressive Integrated Moving Average) model. This allows the trained target model to accurately simulate the dynamic scenes triggered by cultural activities in the target area, improving model performance.

[0046] In some embodiments, the featured scene data mentioned in step 101 further includes featured infrastructure; the featured infrastructure includes narrow dirt roads and temporary construction areas; then step 101 may further include: obtaining basic road data of the target area based on a high-precision map of the target area; extracting images containing narrow dirt roads and temporary construction areas in the target area from UAV aerial photography data based on the basic road data; and performing cubic B-spline curve fitting on the images to obtain the boundaries of the narrow dirt roads and temporary construction areas. In other words, basic road data of the target area can be obtained from a high-precision map provider. High-precision maps provide structured road data, which is suitable for standardized road scenes but may have coverage blind spots, such as narrow dirt roads and temporary construction areas. UAVs can capture dynamic scenes not covered by high-precision maps. Therefore, by comparison, images containing narrow dirt roads and temporary construction areas can be extracted from UAV aerial photography data, and then cubic B-spline curve fitting can be used to generate a smooth, continuous boundary model. This fills the gap in the standardized database regarding featured infrastructure and further improves model performance.

[0047] In some embodiments, the featured scenario data mentioned in step 101 may further include target traffic regulation clauses; then step 101 may further include: using a pre-trained BERT model to extract the target traffic regulation clauses from text data containing local traffic regulations of the target area. That is, specific regulations can be obtained from government documents and legal texts of the target area, and detailed implementation rules can be obtained from traffic management departments. A pre-trained BERT model can be used to extract key regulatory clauses and record them in the text data to train the target large model. This allows the final generated testing strategy to adapt to regional regulatory differences.

[0048] Furthermore, in some embodiments, the method may include: verifying the target test strategy in a simulation environment and obtaining simulation verification metrics; the simulation verification metrics include path planning success rate and emergency braking response time. That is, after the strategy is generated, it can be verified in a simulation environment. The verification metrics include path planning success rate and emergency braking response time. Here, path planning success rate may include the proportion of successfully bypassing temporary obstacles, and emergency braking response time may include the delay from detecting a pedestrian to initiating braking. Thus, through simulation verification, the effectiveness of the strategy can be accurately evaluated, promoting strategy optimization.

[0049] Furthermore, in some embodiments, the process may further include: verifying the target testing strategy in a real vehicle environment to obtain real vehicle verification results; and iteratively optimizing the dynamic scene generator by feeding back the simulation verification metrics and the real vehicle verification results to the target large model. In other words, after completing simulation verification, the strategy is verified in a real vehicle environment. The test vehicle is a real vehicle equipped with an autonomous driving system, with sensor configurations consistent with the data acquisition phase. The acquired data includes sensor data and system decision logs. After completing real vehicle verification, the real vehicle verification results, combined with the preceding simulation verification metrics, are fed back to the model to iteratively optimize the dynamic scene generator. Thus, by optimizing model parameters through simulation and real vehicle verification feedback, the entire process from data acquisition to strategy iteration is automated, effectively improving the reliability of the model.

[0050] In this embodiment, characteristic scene data not covered by the standard scene library in the target area is acquired, including non-standard road markings, traffic participant behavior patterns, and local weather combinations. Then, a target large-scale model is trained based on this characteristic scene data. During training, the target large-scale model learns to extract scene feature vectors from the input data and generates test strategies based on these feature vectors. After training, the model automatically generates target test strategies corresponding to the test scenarios. In this way, by collecting characteristic scene data and constructing a scene library covering regional characteristics, the trained target large-scale model can automatically generate test strategies covering regional characteristics, improving the reliability of test results.

[0051] To provide a more detailed explanation of the solution in this application, a specific embodiment is described below: This embodiment provides an automatic generation scheme for intelligent testing strategies based on large models. The workflow of the system is as follows: Figure 2 As shown, it includes: S21. Collect characteristic scene data in the target area that are not covered by the standard scene library; Specifically, by using multiple sensors and data sources, regionally characteristic scene data is systematically collected to provide structured input for subsequent model training. This step specifically includes the following S211-S217: S211. After completing the basic road data collection, for non-standard road markings in the target area that are not covered by the standard scene library, perform the following operations: Install a high-resolution camera (resolution ≥ 4K, frame rate ≥ 60fps) and a 32-line LiDAR (detection distance ≥ 100m, angular resolution ≤ 0.5°) on the top of the test vehicle; drive in the target area (such as rural roads in Southeast Asia) and collect images and point cloud data including hand-drawn lane lines and temporary traffic signs; For the image data, the adaptive threshold Canny edge detection algorithm is used to extract the outline of the hand-drawn lane lines. The formula is as follows:

[0052] In the formula, For pixels Edge strength at the location; The original image grayscale value; and These are the gradients of the image in the horizontal and vertical directions, respectively; the threshold of this Canny edge detection algorithm is automatically determined by the Otsu algorithm. For point cloud data, the DBSCAN clustering algorithm is used to identify temporary obstacles. The neighborhood radius of this DBSCAN clustering algorithm is... Set to 0.5m, and set the minimum number of points minPts to 5; S212. After completing the collection of non-standard road markings, for local traffic participant behavior patterns in the target area that are not covered by the standard scenario library, perform the following operations: Install 77GHz millimeter-wave radar (detection distance ≥200m, angle accuracy ≤1°) and wide-angle camera (field of view ≥180°) around the test vehicle; drive in the target area (such as urban roads in India) and collect motion trajectory data of disorderly lane changes by motorcycles and pedestrians crossing the road; The extended social force model is used to simulate pedestrian / vehicle motion, and its formula is as follows:

[0053] In the formula, It is the first Each participant at time The total net force acting on it determines its state of motion; It is a goal-oriented force, representing the driving force that enables participants to move toward their destination; It is the sum of interactive forces, representing the other participants. For current participants The resulting social forces (such as avoidance and exclusion); It refers to environmental forces, which represent the constraints imposed on participants by the environment, such as road boundaries and obstacles. S213. After completing the collection of local traffic participant behavior patterns, for dynamic scenarios triggered by cultural activities not covered by the standard scenario library in the target area, perform the following operations: obtain the time and location of cultural activities such as religious festivals and markets from the publicly available calendar of the target area government, and obtain road closure information during the activities from the traffic management department; predict traffic flow changes during the activities based on the ARIMA model, the formula of which is as follows:

[0054] In the formula, It is the traffic flow at the current time t, i.e., the value to be predicted; It is a constant term; It is the autoregressive (AR) coefficient, representing the past... Flow at any given moment For current traffic The extent of the impact; It is the moving average (MA) coefficient, representing the past... Prediction error at each time point Impact on current forecast values; and The order of the model determines how many past moments of data are considered. S214. After completing the dynamic scene collection of cultural activities, for the characteristic infrastructure and temporary obstacles in the target area that are not covered by the standard scene library, perform the following operations: obtain basic road data (accuracy ≤ 10cm) of the target area from a high-precision map provider, and obtain images of narrow dirt roads and temporary construction areas from drone aerial photography data. Based on the acquired imagery, a cubic B-spline curve is used to fit the boundary of the narrow dirt road, and the formula is as follows:

[0055] In the formula, The fitted curve is in the parameters The location coordinates; It is a cubic B-spline basis function, which is a weighting function that determines the degree of influence of each control point on the curve shape; It is the first The coordinates of each control point can be used to change the shape of the curve by adjusting the position of these control points. It refers to the number of control points; S215. After completing the collection of characteristic infrastructure and temporary obstacles, for local weather and environmental interference combinations in the target area that are not covered by the standard scene library, perform the following operations: obtain real-time data such as rainfall, visibility and wind speed from the weather station, obtain data such as road surface humidity and temperature from the vehicle sensor, and generate a rainstorm-visibility combination scene by combining the probability distribution function. S216. After completing the combined data collection of local weather and environmental disturbances, for local special traffic regulations and policies in the target area not covered by the standard scenario library, perform the following operations: obtain special regulations from government documents and legal texts, and obtain the implementation details of the regulations from traffic management departments; use a pre-trained BERT model to extract key clauses of the regulations. The loss function of this BERT model is as follows:

[0056] In the formula, It is the loss function, which measures the difference between the model's prediction and the true label. The training objective is to minimize this loss. This is the total number of training samples; It is the first The true label of each sample; It is the first Text vector representation of each sample; The model predicts that the sample belongs to the true label. The probability of; S217. After completing the data collection for all characteristic scenes, perform the following operations to store the data in a structured manner: Use the LabelMe open-source annotation tool to perform polygon region annotation on the image data (such as hand-drawn lane lines) and cube box annotation on the point cloud data (such as temporary obstacles); use ROS Bag format to store multimodal data, with timestamp synchronization accuracy ≤1ms, and the data packet contains the following fields: data collection timestamp (accuracy ≤1ms), sensor type (such as camera, LiDAR), raw data (such as image pixel values, point cloud coordinates), and annotation information (such as lane line outline, obstacle location). S22. Training cross-domain large models based on collected data; Specifically, the step of constructing a cross-domain large-scale model that can integrate knowledge from multiple domains and generating dynamically combined scenarios includes the following steps S221-S223: S221. After completing data collection and storage, design a cross-domain large-scale model architecture based on the collected characteristic scenario data. It adopts the Transformer Encoder-Decoder structure, with a parameter scale of ≥1 billion, and includes the following modules: Multimodal encoder: processes image, point cloud, and text data respectively, and outputs feature vectors; Cross-modal attention layer: fuses features from different modalities, and its formula is as follows:

[0057]

[0058]

[0059]

[0060] In the formula, and These are feature vectors from different modalities; It is a dimension; (Query) (key), (Value) is determined by their respective weight matrices (i.e. , , The vector obtained after linear transformation of the input features; The function converts similarity scores into weights; the final output is a value vector. The weighted sum is determined by the similarity between the query and the key. This allows the model to focus on relevant parts of one modality based on information from another. Dynamic scene generator: Generates combined scenes based on fusion features; S222. After completing the model architecture design, optimize the model training process based on the collected characteristic scene data: define the scene generation loss by combining the adversarial loss and reconstruction loss of GAN (Generative Adversarial Networks), and use the PPO (Proximal Policy Optimization) algorithm to optimize the model's decision-making ability, whose reward function... The design is as follows:

[0061] S223. After completing model training, deploy the model and establish an update mechanism. The deployment method is to use edge computing nodes (such as NVIDIA Jetson AGX) for real-time inference with a latency of ≤50ms. The input is real-time sensor data, and the output is a dynamic scene feature vector. Based on the federated learning framework, the model parameters of multiple regions are aggregated periodically. S23. Generate intelligent testing strategies; Specifically, based on the trained cross-domain large model, a testing strategy covering regional characteristics is automatically generated. This step includes: After model deployment and updates are completed, test strategies are generated based on the scene feature vectors output by the S22 model. Specifically, when the S22 model outputs a scene feature vector, the system generates the corresponding test strategy through a strategy generator. The scene feature vector includes the following dimensions: road marking type (e.g., hand-drawn lane lines, temporary signs), traffic participant behavior patterns (e.g., motorcycle lane-changing frequency), and environmental interference combination (e.g., heavy rain and visibility). The output test strategy includes the following parameters: sensor calibration frequency (e.g., LiDAR is calibrated once every 100km), path planning optimization step size (e.g., replanning the path every 0.1s), and environmental adaptability test threshold (e.g., triggering emergency mode when visibility is ≤50m during a sandstorm). This policy generator can be considered an optimization model. The system has a built-in rule base that associates specific feature patterns with suggested policy parameters. After generating an initial policy, the policy generator optimizes the policy parameters based on a genetic algorithm. For example, based on all parameters of the test policy corresponding to the scene feature vector, the policy generator randomly generates multiple sets of policy parameters, forming a "population." In a simulation environment (such as CARLA), each set of policy parameters is quickly run, and the optimization is performed based on the fitness function. Calculate the scores, retain the high-scoring policies, and cross-mix and combine the parameters of these high-scoring policies to generate new policies. Apply small random perturbations to some values ​​of the policy parameters, and repeat the evaluation and evolution process until the fitness function is achieved. Convergence, or reaching the preset number of iterations, ultimately yields the optimal combination of test strategy parameters for the current scene's feature vectors; among which, the fitness function... The design can be as follows:

[0062] In the formula, It is the scene coverage rate, which measures how many distinctive scenes a strategy can cover; It refers to decision compliance, which measures whether strategic decisions comply with traffic regulations; It is the system load, used to measure the consumption of computing resources during policy execution; , , These are weighting coefficients used to adjust the relative importance of the three indicators; optionally, , , ; After optimizing the strategy, its effectiveness is verified and fed back to the model. The specific verification metric is the scene coverage rate. and decision compliance Among them, scene coverage This can be the ratio between the number of scenarios actually triggered or covered by the tested strategy and the total number of scenarios in the benchmark scenario library. This benchmark scenario library is built based on all characteristic scenario data collected by S21 and represents all typical situations that need to be tested. When executing the test strategy, the system records which scenarios from the scenario library were encountered during simulation or real-vehicle operation; decision compliance. It can be the ratio between the number of decisions that are deemed compliant and the total number of decisions the system needs to make during the test. Whether a decision is compliant needs to be determined based on preset traffic rules and safety models. Input the validation results into the large model to update the scene generation priority, as shown in the following formula:

[0063] In the formula, and These are the numbers before and after the update. The generation priority of different scenarios; It represents the coverage of this type of scenario under the current testing strategy; It is the learning rate; It is a coverage threshold; optionally, , If coverage Below the threshold If the priority is high, the priority will be lowered; otherwise, it will be higher, thus guiding the model to generate more scenarios with insufficient coverage. S24. Verify the performance of the generated test strategy in both simulation and real vehicle environments; Specifically, after the strategy is generated, the strategy is first verified in a simulation environment using the CARLA open-source autonomous driving simulator. The verification metrics include path planning success rate and emergency braking response time. After completing the simulation verification, the strategy was verified in a real vehicle environment. The test vehicle was equipped with a real vehicle with an L4 level autonomous driving system. The sensor configuration was consistent with the data acquisition phase. The collected data included sensor data (IMU, camera, LiDAR) and system decision logs (such as planned path, braking commands, compliance check records, etc.). After completing the real-vehicle verification, the results are fed back to the model for iterative optimization. In implementation, real-vehicle data is input into the large model to retrain the scene generation module, as shown in the following formula:

[0064] In the formula, and These are the model parameters before and after the update, respectively. It is the total loss function, which includes simulation and real vehicle verification errors; It is the gradient of the total loss function with respect to the model parameters, pointing in the direction in which the loss function grows the fastest; This is the learning rate, a small positive number that controls the step size of each parameter update. By updating the parameters in the opposite direction of the gradient, the loss can be gradually reduced, thus continuously improving the model's performance.

[0065] This embodiment of the solution constructs a scenario library covering regional differences by collecting distinctive data such as non-standard road markings, local traffic behaviors, and dynamic scenes of cultural activities; it integrates knowledge from multiple fields such as traffic engineering, sociology, and meteorology to train a large model that can generate dynamic combination scenarios (such as rainstorm-motorcycle flow-pedestrian mixed traffic), breaking through the single-modal limitation of traditional models; it outputs dynamic generation test strategies based on the large model, and optimizes model parameters through simulation and real vehicle verification feedback, realizing full-process automation from data collection to strategy iteration.

[0066] Corresponding to the embodiments of the aforementioned methods, this application also provides embodiments of a subject positioning guidance device and a terminal for its application: like Figure 3 As shown, Figure 3 This is a block diagram of a subject positioning guidance device provided in an embodiment of this application. The device includes: The acquisition module 31 is used to acquire the current body position of the subject of the capsule endoscope based on sensor data collected by multiple inertial measurement units; the multiple inertial measurement units are arranged on the torso and upper limbs of the subject. The generation module 32 is used to evaluate the difference between the current body position and the target body position, and generate a body position adjustment strategy based on the difference; the body position adjustment strategy is used to guide the subject to adjust the optimal path from the current body position to the target body position. Display module 33 is used to render a target object in real time on the guidance interface based on the body position adjustment strategy; the target object is used to display the position and orientation of different parts of the subject's body and the direction that needs to be adjusted.

[0067] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0068] This application also provides an electronic device, please refer to [link to application]. Figure 4 , Figure 4 This is a structural block diagram of an electronic device provided in an embodiment of this application. The electronic device may include a processor 410, a communication interface 420, a memory 430, and at least one communication bus 440. The communication bus 440 is used to enable direct communication between these components. In this embodiment, the communication interface 420 of the electronic device is used for signaling or data communication with other node devices. The processor 410 may be an integrated circuit chip with signal processing capabilities.

[0069] The processor 410 described above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor, or the processor 410 can be any conventional processor.

[0070] The memory 430 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc. The memory 430 stores computer-readable instructions. When these computer-readable instructions are executed by the processor 410, the electronic device can perform the aforementioned operations. Figure 1 The various steps involved in the method implementation examples.

[0071] Alternatively, the electronic device may also include a storage controller and an input / output unit.

[0072] The memory 430, storage controller, processor 410, peripheral interface, and input / output unit are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses 440. The processor 410 is used to execute executable modules stored in the memory 430, such as software function modules or computer programs included in electronic devices.

[0073] The input / output unit is used to provide users with the ability to create tasks and to set optional start periods or preset execution times for those tasks, thereby enabling user-server interaction. The input / output unit may be, but is not limited to, a mouse and keyboard.

[0074] Understandable. Figure 4 The structure shown is for illustrative purposes only; the electronic device may also include components that are more advanced than those shown. Figure 4 The more or fewer components shown, or having the same Figure 4 The different configurations shown. Figure 4 The components shown can be implemented using hardware, software, or a combination thereof.

[0075] This application also provides a storage medium storing instructions. When the instructions are run on a computer, the computer program is executed by a processor to implement the method described in the method embodiment. To avoid repetition, the method will not be described again here.

[0076] This application also provides a computer program product that, when run on a computer, causes the computer to perform the method described in the method embodiment.

[0077] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0078] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0079] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0080] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0082] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for generating vehicle testing strategies, characterized in that, include: Acquire distinctive scene data in the target area that is not covered by the standard scene library; The featured scenario data includes non-standard road markings, traffic participant behavior patterns, and local weather combinations; The target large model is trained based on the aforementioned characteristic scene data to obtain a trained target large model; during the training process, the target large model learns to extract scene feature vectors from the input data and then generates a test strategy based on the scene feature vectors. Based on the trained target model, a target testing strategy corresponding to the test scenario is generated.

2. The method according to claim 1, characterized in that, The non-standard road markings include hand-drawn lane lines and temporary obstacles; The acquisition of distinctive scene data in the target area that is not covered by the standard scene library includes: Image data and point cloud data containing non-standard road markings were collected by test vehicles traveling in the target area; The Canny algorithm with adaptive thresholding is used to extract the hand-drawn lane lines from the image data, and the image data is then labeled based on the hand-drawn lane lines. The DBSCAN clustering algorithm is used to identify the temporary obstacles from the point cloud data, and the point cloud data is labeled according to the temporary obstacles.

3. The method according to claim 2, characterized in that, The acquisition of distinctive scene data in the target area that is not covered by the standard scene library also includes: Non-standard traffic motion trajectory data were collected by test vehicles traveling in the target area; Based on the non-standard traffic trajectory data, an extended social force model is used to simulate the movement of pedestrians and vehicles, thereby obtaining the behavior patterns of the traffic participants. The behavioral patterns of the traffic participants are recorded in text data.

4. The method according to claim 3, characterized in that, The acquisition of distinctive scene data in the target area that is not covered by the standard scene library also includes: Real-time meteorological data for the target area is obtained from a weather station; the real-time meteorological data includes rainfall, visibility, and wind speed. Environmental data is acquired from the onboard sensors of the test vehicle located in the target area; The local weather combination is obtained by combining the real-time meteorological data and the environmental data; The local weather combinations are recorded in the text data.

5. The method according to claim 4, characterized in that, The target large model includes a combined scenario generation sub-model and a test strategy generation sub-model; The combined scene generation sub-model includes a multimodal encoder, a cross-modal attention layer, and a dynamic scene generator; the multimodal encoder is used to process the input image data, point cloud data, and text data respectively, and output the corresponding feature vectors; the cross-modal attention layer is used to fuse the feature vectors of different modalities to generate fused features; The dynamic scene generator is used to generate a combined scene based on the fusion features; The test strategy generation sub-model is used to generate a test strategy based on the scene feature vector corresponding to the combined scene; The testing strategy includes the following parameters: sensor calibration frequency, path planning optimization step size, and environmental adaptability test threshold.

6. The method according to claim 1, characterized in that, The featured scene data also includes dynamic scene data triggered by cultural activities; the dynamic scene data includes changes in traffic flow. The acquisition of distinctive scene data in the target area that is not covered by the standard scene library includes: Obtain the time and location of cultural events in the target area; Based on the time and location, obtain road closure information during the event; Using the road closure information, traffic flow changes during the event are predicted based on the ARIMA model.

7. The method according to claim 1, characterized in that, The featured scenario data also includes featured infrastructure; the featured infrastructure includes narrow dirt roads and temporary construction areas. The acquisition of distinctive scene data in the target area that is not covered by the standard scene library includes: Based on the high-precision map of the target area, obtain the basic road data of the target area; Based on the basic road data, images containing narrow dirt roads and temporary construction areas in the target area are extracted from the drone aerial photography data. The boundary of the narrow dirt road and the temporary construction area is obtained by performing cubic B-spline curve fitting on the image.

8. The method according to claim 1, characterized in that, The specific scenario data also includes target traffic regulations and provisions; The acquisition of distinctive scene data in the target area that is not covered by the standard scene library includes: Using a pre-trained BERT model, the target traffic regulation clauses are extracted from text data containing local traffic regulations for the target area.

9. The method according to claim 5, characterized in that, Also includes: The target testing strategy is verified in a simulation environment to obtain simulation verification metrics; the simulation verification metrics include path planning success rate and emergency braking response time.

10. The method according to claim 9, characterized in that, Also includes: The target testing strategy was verified in a real vehicle environment to obtain real vehicle verification results. The dynamic scene generator is iteratively optimized by feeding back the simulation verification metrics and the real vehicle verification results to the target large model.

11. A vehicle testing strategy generation device, characterized in that, include: The acquisition module is used to acquire characteristic scene data in the target area that is not covered by the standard scene library; The featured scenario data includes non-standard road markings, traffic participant behavior patterns, and local weather combinations; The training module is used to train a target large model based on the aforementioned characteristic scene data, and obtain a trained target large model. During the training process, the target large model learns to extract scene feature vectors from the input data and then generates a test strategy based on the scene feature vectors. The generation module is used to generate a target testing strategy corresponding to the test scenario based on the trained target model.

12. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 10.

13. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any one of claims 1 to 10.

Images