An automatic driving test scene library construction method based on real traffic data
By selective sensor fusion and deep learning models based on real traffic data, combined with association rule mining technology, an autonomous driving test scenario library is constructed. This solves the problems of low automation and ambiguous risk assessment in existing technologies, achieving more accurate scenario recognition and risk classification, and improving the test coverage and safety of autonomous driving systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGAN UNIV
- Filing Date
- 2025-06-03
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for constructing autonomous driving test scenario libraries rely on manual annotation, resulting in low automation, vague risk assessment indicators, and difficulty in fine-grained classification. Consequently, the test scenario libraries have limited capabilities in risk assessment and management, making it difficult to support diverse testing needs.
By employing a method based on real traffic data, a selective sensor fusion framework and a deep learning model are used to identify scene elements and construct a risk assessment model. Combined with association rule mining technology, an autonomous driving test scenario library is built to achieve automated classification and diversity of scene risks.
It improves the diversity and relevance of test scenarios, enhances the safety and stability of autonomous driving systems, supports the rapid retrieval of test scenarios with specific features and risk levels, and promotes the commercialization of autonomous driving technology.
Smart Images
Figure CN120631991B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving technology, specifically a method for constructing an autonomous driving test scenario library based on real traffic data. Background Technology
[0002] With the rapid development of autonomous driving technology, its commercialization has made tremendous progress. To accelerate the deployment of autonomous driving technology, it is necessary to construct autonomous driving test scenarios to extract diverse test cases for comprehensive testing of autonomous driving systems. However, existing methods for constructing test scenario libraries have some shortcomings. These methods largely rely on manual annotation, resulting in low automation levels. Furthermore, risk assessment indicators are vague and difficult to quantify, and there is a lack of refined classification of scenario elements and risk types. This limits the risk assessment and management capabilities of the test scenario library, making it difficult to effectively support the diverse testing needs of autonomous driving systems. Summary of the Invention
[0003] The purpose of this invention is to provide a method for constructing an autonomous driving test scenario library based on real traffic data. By comprehensively capturing the characteristics of traffic scenarios, identifying scenario elements, and scientifically classifying the risk levels of scenarios, a scenario library can be constructed, thereby improving the diversity and relevance of test scenarios.
[0004] To achieve the above objectives, this invention provides a method for constructing an autonomous driving test scenario library based on real traffic data, comprising the following steps:
[0005] Step 1. Based on the selective sensor fusion framework, dynamically adjust the fusion strategy of multimodal sensors according to the current driving environment. The selective sensor fusion framework is configured as follows:
[0006] The system is equipped with multiple convolutional neural networks (CNNs). Each input sensor data is fed into its corresponding CNN to extract data features F. i According to the data feature F i The system identifies the context information of the current scene and selects a sensor fusion strategy suitable for the current scene based on the context information; it then uses a preset deep learning model branch to process the corresponding sensor fusion strategy to obtain the target scene result and the corresponding scene elements.
[0007] Step 2. Standardize the scene elements obtained in Step 1 by unifying their naming and data formats, then classify them into layers, select key elements in each layer of the scene as risk assessment indicators, construct a risk assessment model, calculate the comprehensive risk score, and initially classify the risk levels.
[0008] Step 3. Based on the scene element hierarchical classification results and preliminary risk level classification results in Step 2, the association rule mining technology is used to mine the association rules between the scene elements and the risk level, and a rule-based classifier is constructed to classify the risk level of the scene to be evaluated.
[0009] Step 4. Structure and store scenario elements and risk levels to build an autonomous driving scenario library that supports multi-dimensional queries. This library can flexibly select test sets with specific elements and risk levels according to different testing needs.
[0010] The deep learning model has multiple branches, and each deep learning model branch processes a specific sensor fusion strategy.
[0011] The processing results of the deep learning model branch on the sensor fusion strategy include at least the target's bounding box, confidence score, and type label.
[0012] It also includes performing nonmaximum suppression or weighted box fusion on the processing results of the deep learning model branch on the sensor fusion strategy.
[0013] Step 2, the scene elements are divided into six layers of scene elements:
[0014] The main vehicle level includes the status of the vehicle being tested;
[0015] The weather layer includes external environmental conditions that affect perception and vehicle dynamics;
[0016] Traffic participant layer, including dynamic obstacle types and behavioral intentions;
[0017] Temporary construction layer, including non-standard road modification elements;
[0018] The traffic facilities layer includes fixed traffic management and guidance facilities;
[0019] The road topology layer includes the structured features of the road network.
[0020] The calculation formula for constructing the risk assessment model described in step 2 is as follows:
[0021]
[0022] Where R represents the comprehensive risk score, x i For the score of risk assessment indicator i, w i Let be the weight of risk assessment indicator i, and n represent the number of risk assessment indicators.
[0023] Based on the percentage range of scenario risk, scenarios are divided into three risk levels: normal, critical, and dangerous. The risk levels are divided by percentage: normal (0% ≤ R < 30%), critical (30% ≤ R < 70%), and dangerous (R ≥ 70%).
[0024] Scenario risk percentage R ％ The calculation formula is as follows:
[0025]
[0026] Where R min R represents the lowest risk score. max This indicates the highest risk score.
[0027] In step 3, the association rule mining technique is the Apriori algorithm;
[0028] Through the layer-by-layer iterative mechanism of the Apriori algorithm, the association between risk level and combination of multiple elements from a single element is discovered, forming the original rule base of element combination-risk level.
[0029] The generated original rule base is cleaned and deduplicated. Weights are assigned to all cleaned rules to obtain a rule-based classifier. For each scenario to be evaluated, the scenario elements are matched with the rules in the classifier. The weight of each successfully matched rule is calculated, and a comprehensive score is determined for the risk level of the scenario. The calculation formula is as follows:
[0030]
[0031] Where n is the number of matching rules, X i Y represents the risk level, where i is the score of the risk assessment indicator and Y is the risk level. An aggressive strategy is used to classify the risk levels.
[0032] Using an aggressive strategy for risk level classification means:
[0033] The risk level with the highest comprehensive score will be used as the final classification result;
[0034] If the calculated high-risk comprehensive score equals the low-risk comprehensive score, then it is classified as high-risk.
[0035] The multimodal sensor includes at least a camera and a lidar.
[0036] A computer-readable storage medium storing computer program instructions, which can be executed by a processor to implement the above-described method for constructing an autonomous driving test scenario library based on real traffic data.
[0037] (iii) Beneficial effects:
[0038] Compared with existing technologies, this method for constructing an autonomous driving test scenario library based on real traffic data has the following advantages:
[0039] I. This invention processes multimodal sensor data collected in traffic scenarios to achieve feature recognition and element extraction. Combined with a hierarchical structure analysis method for scene elements, a risk assessment model is constructed to achieve preliminary risk level classification. An association rule mining algorithm is used to obtain association rules between scene elements and risk levels, and a rule-based scene classifier is constructed to automate scene risk classification. Finally, an autonomous driving test scenario library based on scene elements and risk levels is built. This method can more comprehensively capture the characteristics of traffic scenarios, more accurately identify scene elements, and scientifically classify scene risk levels. Combining these two aspects to construct the scenario library improves the diversity and relevance of test scenarios, providing stronger support for the development and verification of autonomous driving technology.
[0040] Second, the construction method provided by this invention not only improves the coverage of test scenarios but also enables rapid retrieval of test scenarios with specific characteristics and risk levels based on different testing needs. This will help improve the safety and stability of autonomous driving systems and accelerate their commercialization process. Attached Figure Description
[0041] Figure 1 This is a schematic diagram of the sensor data fusion and feature recognition process provided in an embodiment of the present invention;
[0042] Figure 2 This is a schematic diagram of the overall method for constructing the scene library of the present invention;
[0043] Figure 3 This is a schematic diagram of the overall method for constructing the scene library of the present invention; Detailed Implementation
[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0045] like Figure 1-3 As shown, the present invention provides a technical solution: a method for constructing an autonomous driving test scenario library based on real traffic data, comprising the following steps:
[0046] Step 1. Based on the selective sensor fusion framework, dynamically adjust the fusion strategy of multimodal sensors according to the current driving environment. The selective sensor fusion framework is configured as follows:
[0047] The system is equipped with multiple convolutional neural networks (CNNs). Each input sensor data is fed into its corresponding CNN to extract data features F. i According to the data feature F i The system identifies the context information of the current scene and selects a sensor fusion strategy suitable for the current scene based on the context information; it then uses a preset deep learning model branch to process the corresponding sensor fusion strategy to obtain the target scene result and the corresponding scene elements.
[0048] Autonomous driving systems are equipped with multiple sensors to collect environmental information. Existing sensor fusion technologies, mostly employing static fusion strategies, have limitations in complex scenarios. For example, pre-fusion is prone to failure when sensor data is noisy or obstructed, and post-fusion cannot fully utilize multi-sensor features, leading to decreased estimation accuracy. To improve data fusion and feature recognition rate, this invention uses the HydraFusion selective sensor fusion framework to form a multimodal sensing fusion strategy. It learns and recognizes contextual information of the driving scenario (such as weather, lighting, road type, etc.) and selects the optimal sensor combination for fusion based on this information, thereby improving perception robustness without increasing computational complexity and energy consumption. Its workflow diagram is shown below. Figure 1 As shown.
[0049] HydraFusion accepts input data from multiple sensors, with each sensor's data fed into its corresponding convolutional neural network (CNN), called a "Stem". The Stem transforms the sensor data into a feature representation for subsequent context recognition and fusion strategy selection. For each sensor S... i Its corresponding Stem i It will extract data features F i :
[0050] F i =Stem i (S i )
[0051] i represents the sensor (such as a camera and LiDAR), and the features of all sensors are combined into a data feature set F:
[0052] F = {F radar ,F camera …}
[0053] The features from all sensors are combined into a single data feature set F, and the feature set of the lidar is F. radar The feature set of the camera is F camera ;
[0054] The system identifies contextual information about the current scene based on features, including light intensity and road type. This contextual information is then output by the context recognition model. The gating module dynamically selects the optimal sensor fusion strategy for the current scene based on this information; that is, it determines which sensor data should be fused and how to fuse them, based on the current context.
[0055] Based on the optimal sensor fusion strategy determined by the gating module, a set of selected deep learning model branches are executed. Each branch model processes a specific combination of sensor data (the branch can be single sensor data or a multi-sensor fusion branch) and generates detection results. Radar branch detection is performed on a scene. radar Fusion detection of camera and LiDAR radar+camera :
[0056] O radar = radar (F radar )
[0057] O radar+camera = radar+camera (F radar +F camera )
[0058] The output of each branch is O * It is a set of object detection results, including information such as the object's bounding box, confidence score, and type label. By using non-maximum suppression or weighted box fusion (WBF), the optimal detection result is selected to improve the accuracy and robustness of object detection.
[0059] The formula for using the WBF algorithm is as follows:
[0060]
[0061] For each scene, there are m branches of detection results. The coordinates of the target bounding box detected by the k-th branch are (μ k,1 v k,1 μ k,2 v k,2 ), with a confidence level of C k . and This represents the coordinates of the final fused bounding box, ultimately yielding the detection result for each target: O = Fusion(all O...). * ).
[0062] This framework outputs a final list of detected targets for each scene (target category, quantity, etc.), as well as identified scene context information (weather type, lighting conditions, road type, etc.). This information, which reflects scene characteristics, state, and dynamic changes, is called scene elements. After extracting elements for a single scene, the target categories and context information from all scenes are aggregated and standardized through data cleaning, naming conventions, and data format standardization to ensure data accuracy and consistency, forming macroscopic traffic scene elements for subsequent analysis.
[0063] Based on the standardized results, all data for each scenario are organized to ensure the integrity and consistency of the information of each scenario's elements, and each scenario and its element information are organized into a structured format such as JSON or XML.
[0064] Step 2. Standardize the scene elements obtained in Step 1 by unifying their names and data formats, then classify them into layers, select key elements in each layer of the scene as risk assessment indicators, construct a risk assessment model, calculate the comprehensive risk score, and initially classify the risk levels.
[0065] A six-layer scenario element classification framework is adopted, using a structured hierarchical method to systematically decouple and classify the macroscopic traffic scenario elements summarized in Step 1. The six layers encompass a complete set of scenario elements at six levels: the vehicle layer, the weather layer, the traffic participants layer, the temporary construction layer, the traffic facility layer, and the road topology layer. Each layer independently represents specific dimensions of risk characteristics, and through interactive relationships (such as the coupling influence of the weather layer on the perceived visibility of the traffic participants layer), they form the basis for systematic risk analysis.
[0066] The six layers of scene elements are categorized as follows:
[0067] Main vehicle layer: This layer covers the status of the vehicle being tested, such as speed and yaw angle.
[0068] Weather layer: External environmental conditions that affect perception and vehicle dynamics, such as sunlight, precipitation, and wind speed.
[0069] Traffic participant layer: dynamic obstacles and their behavioral intentions, such as motor vehicles and pedestrians.
[0070] Temporary construction layer: Non-standardized temporary road change elements, such as guide gestures, temporary traffic lights, etc.
[0071] Traffic facilities layer: Fixed traffic management and guidance facilities, such as traffic light phases, speed limit signs, and solid / dual lane markings.
[0072] Road topology layer: The structural characteristics of the road network, such as the number of lanes, ramp connections, intersection types, etc.
[0073] From all scenario elements, several key elements that have a significant impact on risk assessment are selected as assessment indicators and assigned weight coefficients to classify risk levels.
[0074] A risk assessment model is constructed by setting scoring criteria and corresponding weighting coefficients for different indicators such as vehicle status data (vehicle speed, collision time, etc.) and environmental data (traffic flow, road type, weather conditions, etc.):
[0075]
[0076] Where R represents the comprehensive risk score, x i For the score of indicator i, w i Let be the weight of indicator i, and n represent the number of indicators. Based on the risk percentage range, scenarios are divided into three risk levels: normal, critical, and dangerous. Scenario risk percentage R ％ The calculation formula is as follows:
[0077]
[0078] Where R min R represents the lowest risk score. max This indicates the highest risk score.
[0079] Table 1. Risk Classification
[0080]
[0081]
[0082] In step 1, the element information for each traffic scenario has been extracted and standardized, and organized into a structured format. In step 2, based on the constructed risk assessment model, the key risk elements of each traffic scenario are quantitatively analyzed to obtain their risk levels, and these risk levels are incorporated into the structured description of each scenario.
[0083] Step 3. Based on the scene element hierarchical classification results and preliminary risk level classification results in Step 2, the association rule mining technology is used to mine the association rules between the scene elements and the risk level, and a rule-based classifier is constructed to classify the risk level of the scene to be evaluated.
[0084] The association rule mining technique used is the Apriori algorithm;
[0085] Through the layer-by-layer iterative mechanism of the Apriori algorithm, the association between risk level and combination of multiple elements from a single element is discovered, forming the original rule base of element combination-risk level.
[0086] The Apriori algorithm is used to calculate support, confidence, and lift, and rules that meet the threshold are then selected.
[0087] Among them support The calculation formula is as follows:
[0088]
[0089] Where X represents traffic elements, Y represents risk level, P(X∩Y) represents the probability of having both X and Y, N(X∩Y) is the number of scenarios that simultaneously contain X and Y, and N represents the total number of scenarios;
[0090] Support measures the proportion of a certain type of traffic element and a specific risk level occurring across all scenarios. By setting a support threshold, rules with high predictive power are selected.
[0091] Confidence This represents the probability of scenario Y occurring in a scenario containing certain specific elements X. The calculation formula is as follows:
[0092]
[0093] Where P(Y|X) represents the probability of Y occurring given that X occurs, and Support(X) represents the proportion of scenarios where traffic element X occurs to the total number of scenarios. Confidence level is used to assess the probability of a certain type of risk occurring when certain traffic elements occur. By setting a minimum confidence threshold, high-frequency and statistically significant scenarios are retained, avoiding noise interference.
[0094] Improvement The formula used to measure the strength of the correlation between elements X and Y (more valuable than random guessing) is as follows:
[0095]
[0096] Where P(X) represents the probability of X occurring independently, P(Y) represents the probability of Y occurring independently, Support(X) represents the proportion of scenarios in which traffic element X occurs to the total number of scenarios, and Support(Y) represents the proportion of scenarios in which risk level Y occurs to the total number of scenarios. Lift is used to reveal the dependency relationship between the combination of elements and the risk level, that is, whether the occurrence of a specific combination of elements significantly affects the probability of occurrence of a certain risk level, so as to identify rules that are truly causal or strongly correlated (Lift>1 indicates positive correlation, Lift=1 indicates independence, and Lift<1 indicates negative correlation).
[0097] Through the iterative mechanism of the Apriori algorithm, the correlation between risk level and single factor combinations is discovered, forming (factor combination X...). * The original rule base for risk level Y;
[0098] The generated original rule base is cleaned and deduplicated. Weights are assigned to all cleaned rules, and a rule-based classifier is constructed. For each scene feature to be classified, it is matched against the rules in the classifier. The weight of each successfully matched rule is calculated, and a comprehensive score is calculated for each risk level. The calculation formula is as follows:
[0099]
[0100] Where n is the number of matching rules, X * i It is the i-th feature combination, Y is the risk level, and an aggressive strategy is used to classify the risk level.
[0101] Using an aggressive strategy for risk level classification means:
[0102] The risk level with the highest comprehensive score will be used as the final classification result;
[0103] If the calculated high-risk comprehensive score equals the low-risk comprehensive score, then it is classified as high-risk.
[0104] The cleaning and de-duplication process is as follows:
[0105] Merge duplicate rules: If two rules have the same content, keep the one with the higher confidence level.
[0106] Eliminate redundant rules: If the preconditions of rule A are a subset of the preconditions of rule B, and both rules have the same conclusion, delete the rule with the smaller coverage.
[0107] Conflict resolution: If two rules have the same preconditions but different conclusions, retain the rule with higher confidence.
[0108] Step 4. Structure and store scenario elements and risk levels to build an autonomous driving scenario library that supports multi-dimensional queries. Based on the constructed autonomous driving scenario library, those skilled in the art can flexibly select test sets with specific elements and risk levels according to different testing needs, forming corresponding autonomous driving test scenarios.
[0109] A data model is constructed to organize and store the multi-layered elements and risk levels of each scenario, enabling complex queries and multi-dimensional analysis. The organization and storage method should support complex queries based on single elements or combinations of multiple elements, quickly locating specific test scenarios and supporting analysis from different dimensions (such as risk level, scenario element type, etc.) to meet the complex needs of autonomous driving system testing. It should be able to flexibly select test sets with specific elements and risk levels according to different testing requirements, improving the targeting and efficiency of testing. For example, testers can select scenarios containing specific elements (such as highways, night driving, and foggy weather) with a medium risk level from numerous traffic scenarios for testing.
[0110] A computer-readable storage medium storing computer program instructions, which can be executed by a processor to implement the above-described method for constructing an autonomous driving test scenario library based on real traffic data.
[0111] Here is a suggested organizational storage method:
[0112] Scenario: Scenario number, risk level label, and scenario storage path.
[0113] Main vehicle: speed, yaw angle, vehicle type, etc.
[0114] Weather layer: visibility, rainfall, wind speed, etc.
[0115] Traffic participants layer: type, quantity, etc.
[0116] Temporary construction layer: construction signs, construction type, etc.
[0117] Transportation facilities level: facility status, type, quantity, etc.
[0118] Road topology layer: road width, curvature, type, slope, etc.
[0119] The implementation principle of this invention, a method for constructing an autonomous driving test scenario library based on real traffic data, is as follows: Multimodal sensor data collected from traffic scenarios is processed to achieve feature recognition and element extraction; a risk assessment model is constructed using a hierarchical structure analysis method for scenario elements to achieve preliminary risk level classification; association rule mining algorithms are used to obtain association rules between scenario elements and risk levels, and a rule-based scenario classifier is constructed to automate scenario risk classification; finally, an autonomous driving test scenario library based on scenario elements and risk levels is built. This invention can more comprehensively capture the characteristics of traffic scenarios, more accurately identify scenario elements, and scientifically classify the risk levels of scenarios. By combining these two aspects to construct the scenario library, the diversity and relevance of test scenarios are improved, providing stronger support for the development and verification of autonomous driving technology.
[0120] This construction method not only improves the coverage of test scenarios but also enables the rapid retrieval of test scenarios with specific characteristics and risk levels based on different testing needs. This will help improve the safety and stability of autonomous driving systems and accelerate their commercialization process.
Claims
1. A method for constructing an autonomous driving test scenario library based on real traffic data, characterized in that, Includes the following steps: Step 1. Based on the selective sensor fusion framework, dynamically adjust the fusion strategy of multimodal sensors according to the current driving environment. The selective sensor fusion framework is configured as follows: The system is equipped with multiple convolutional neural networks (CNNs). Each sensor data input is fed into its corresponding CNN to extract data features. Based on the data characteristics The system identifies the context information of the current scene and selects a sensor fusion strategy suitable for the current scene based on the context information; it then uses a preset deep learning model branch to process the corresponding sensor fusion strategy to obtain the target scene result and the corresponding scene elements. Step 2. Standardize the information related to the scene elements obtained in Step 1, then classify the scene elements into layers, select key elements in each layer of the scene as risk assessment indicators, construct a risk assessment model, calculate the comprehensive risk score, and initially classify the risk level. Step 3. Based on the scene element hierarchical classification results and preliminary risk level classification results in Step 2, the association rule mining technology is used to mine the association rules between the scene elements and the risk level, and a rule-based classifier is constructed to classify the risk level of the scene to be evaluated. Step 4. Structure and store scenario elements and risk levels to build an autonomous driving scenario library that supports multi-dimensional queries; In step 3, the association rule mining technique is the Apriori algorithm; Through the layer-by-layer iterative mechanism of the Apriori algorithm, the association between risk level and combination of multiple elements from a single element is discovered, forming the original rule base of element combination-risk level. The generated original rule base is cleaned and deduplicated, and all cleaned rules are assigned weights to obtain a rule-based classifier. For each scenario to be evaluated, the scenario elements are matched with the rules in the classifier. For each successfully matched rule, its weight is calculated, and a comprehensive score is calculated for the risk level of the scenario. The calculation formula is as follows: ； in For the number of matching rules, Risk assessment indicators The score, It is a risk level classification, and an aggressive strategy is used to classify risk levels.
2. The method for constructing an autonomous driving test scenario library based on real traffic data according to claim 1, characterized in that: The deep learning model has multiple branches, and each deep learning model branch processes a specific sensor fusion strategy. The processing results of the deep learning model branch on the sensor fusion strategy include at least the target's bounding box, confidence score, and type label.
3. The method for constructing an autonomous driving test scenario library based on real traffic data according to claim 1, characterized in that: It also includes performing nonmaximum suppression or weighted box fusion on the processing results of the deep learning model branch on the sensor fusion strategy.
4. The method for constructing an autonomous driving test scenario library based on real traffic data according to claim 1, characterized in that: Step 2, the scene elements are divided into six layers of scene elements: The main vehicle level includes the status of the vehicle being tested; The weather layer includes external environmental conditions that affect perception and vehicle dynamics; Traffic participant layer, including dynamic obstacle types and behavioral intentions; Temporary construction layer, including non-standard road modification elements; The traffic facilities layer includes fixed traffic management and guidance facilities; The road topology layer includes the structured features of the road network.
5. The method for constructing an autonomous driving test scenario library based on real traffic data according to claim 1, characterized in that: The calculation formula for constructing the risk assessment model described in step 2 is as follows: ； in, This represents the overall risk score. Risk assessment indicators The score, Risk assessment indicators The weight, Indicates the number of risk assessment indicators; Based on the percentage range of scenario risk, scenarios are divided into three risk levels: normal, critical, and dangerous. The risk levels are divided by percentage: normal is 0% ≤ R < 30%, critical is 30% ≤ R < 70%, and dangerous is R ≥ 70%. Percentage of scenario risks The calculation formula is as follows: ； in Indicates the lowest risk score. This indicates the highest risk score.
6. The method for constructing an autonomous driving test scenario library based on real traffic data according to claim 1, characterized in that: The adoption of an aggressive strategy for risk level classification refers to: The risk level with the highest comprehensive score will be used as the final classification result; If the calculated high-risk comprehensive score equals the low-risk comprehensive score, then it is classified as high-risk.
7. The method for constructing an autonomous driving test scenario library based on real traffic data according to claim 1, characterized in that: The multimodal sensor includes at least a camera and a lidar.
8. A computer-readable storage medium, characterized in that, The system stores computer program instructions that can be executed by a processor to implement the method for constructing an autonomous driving test scenario library based on real traffic data as described in any one of claims 1-7.