Intelligent driving system test scene building system based on natural driving environment collection data

Abstract

The invention relates to the technical field of automatic driving, in particular to an intelligent driving system test scene building system based on natural driving environment collection data. The system sequentially comprises data preprocessing and quality control, automatic screening of a test main vehicle, automatic identification of a high-value test scene based on the main vehicle, similarity measurement of the test scene, clustering based on the similarity measurement of the test scene and representative simulation test scene reconstruction extracted based on a clustering result. According to the method, by means of noise deletion, track smoothing and the like, the quality of original data is remarkably improved, the influence of noise data is reduced, the accuracy and reliability of subsequent analysis are improved, and a set of comprehensive test scene recognition rules is utilized to automatically recognize a high-value test scene.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and specifically to a system for building intelligent driving system test scenarios based on data collected in a natural driving environment. Background Art

[0002] With the rapid development of autonomous driving technology, the testing and verification of autonomous vehicles has become increasingly important. In order to establish a complete autonomous vehicle verification platform, the International Automobile Manufacturers Association has proposed a "three-pillar" testing and certification method for autonomous vehicles, consisting of virtual simulation testing, closed-field testing, and actual road testing. Among them, virtual simulation testing will dominate the testing of autonomous vehicles due to its high efficiency, low cost, process safety, and reproducible scenarios.

[0003] Reconstructing real-world test scenarios in a simulation environment is one of the important means of virtual simulation testing. However, how to construct test data in a simulation environment to ensure its authenticity, effectiveness and efficiency has always been a key issue in simulation testing. In order to meet the needs of intelligent vehicle simulation testing, especially in the case of rapid algorithm updates and iterations, there is an urgent need for a method that can efficiently search and reconstruct high-authenticity, high-value test scenarios from natural driving data collected in the real world.

[0004] Early smart car testing scenarios mainly came from traffic accident data from the transportation department, and a corresponding test scenario library was generated through restoration and reconstruction. With the development of science and technology, the development of various advanced sensors and natural driving data acquisition technology has made significant progress, but there are still the following problems in the identification and reconstruction of test scenarios: 1. The proportion of noise data and erroneous data in the original collected data is high: Due to the complexity of the acquisition environment and the limitations of the sensors, the original data inevitably contains a large amount of noise data and erroneous data, which directly affects the accuracy and efficiency of subsequent data processing; 2. The identification process of test scenarios depends on engineer experience and manual operation: The current method is highly dependent on the professional knowledge and experience of engineers for data screening and scene identification, which not only increases the workload, but also may lead to the omission or misjudgment of important information; 3. The identification of test scenarios is highly bound to the test conditions: Existing methods are often limited to the identification of known conditions, and are limited by the thinking logic of engineers, and cannot effectively discover unknown conditions or unsafe scenarios; 4. The extracted test scenarios have a high degree of similarity, and the effective test ratio of the system to be tested is low: Since existing methods often focus on common or easy-to-identify scenarios, the extracted test scenarios often show a high degree of similarity, which leads to a relatively low proportion of effective testing of intelligent driving systems, and it is difficult to fully cover all potential dangerous conditions. Summary of the invention

[0005] The purpose of the present invention is to provide a system for building intelligent driving system test scenarios based on data collected in a natural driving environment, so as to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a system for building test scenarios for an intelligent driving system based on data collected from a natural driving environment, the system sequentially including data preprocessing and quality control, automatic screening of a test vehicle, automatic identification of high-value test scenarios based on the vehicle, similarity measurement of the test scenarios, clustering based on the similarity measurement of the test scenarios, and reconstruction of representative simulation test scenarios based on the extraction of clustering results; The data preprocessing and quality control include trajectory noise point deletion, noise target deletion, trajectory smoothing and target merging; the trajectory noise point deletion introduces the DBSCAN clustering algorithm to determine whether the trajectory is noise data according to the degree of aggregation of trajectory points, delete the noise trajectory, and only retain one trajectory point in the vehicle stationary stage to reduce redundancy; the noise target deletion means that if the existence time of the target is less than 3 seconds, the target is considered to be noise data and deleted; the trajectory smoothing uses the cubic spline interpolation method to deal with discontinuities or mutations in the collected data to obtain a smoother and more coherent vehicle trajectory; The test vehicle is automatically screened according to the following three rules: Rule 1: The main vehicle must travel at least 100 meters; Rule 2: The main vehicle must continue to move for at least 15 seconds; Rule 3: The proportion of the main vehicle’s non-stationary time to the total time is at least 50%; If the above three rules are met at the same time, it can be used as the main vehicle; The automatic identification of high-value test scenarios based on the main vehicle is as follows: based on each of the above-screened main vehicles, seven screening rules are used to automatically identify high-value test scenarios. The seven screening rules include: Rule 1: Identification of the risk scenario of a forward collision with the main vehicle; Rule 2: Identify pedestrians or large vehicles in the front at a close distance; Rule 3: Intersection detection of the tracks of the host vehicle and the target vehicle; Rule 4: Identification of scenarios where traffic efficiency is limited; Rule 5: Identification of congested driving areas; Rule 6: Identification of driving area restricted scenes; Rule 7: Identify the rapid changes in the relative state of the target object; Any one or more of the above rules can be used as the basic test scenario; The similarity measurement of the test scene is to measure the similarity of the basic test scene attributes obtained above by six scene attributes, wherein the six scene attributes include: terrain characteristics, main vehicle trajectory shape, target object trajectory shape, conflict type characteristics, target object type characteristics and external input conditions; The clustering based on the test scene similarity measurement can calculate the similarity between two targets based on the above six scene attributes, thereby obtaining a similarity matrix. Combined with the neighbor propagation clustering analysis algorithm, the test scenes can be clustered and analyzed. By calculating the similarity between the scenes, scene groups with common characteristics can be identified, and the cluster center scenes can be extracted as high-value test scenes. The representative simulation test scenario reconstruction extracted based on clustering results is based on representative high-value test scenarios obtained by clustering based on test scenario similarity metrics, and its original data is extracted, and the trajectory and speed data of all traffic participants are automatically converted into scenario files that can be directly used in the simulation environment, reproducing the movement trajectory and behavior of traffic participants in the simulation environment.

[0007] Preferably, the data preprocessing and target merging in quality control comprises the following steps: Step 1, read the original data and the smoothed data; Step 2, extracting the status data of each target object; Step 3: Analyze each target object and record the start and end states; Step 4: traverse all targets and predict the possible position range of the target within 5 seconds based on the target's end state. If the end state of another target coincides with this range, a merged map is constructed. Step five: process the data of the target object in the data according to the merging relationship.

[0008] Preferably, the status data includes timestamp, speed, acceleration, position, heading angle and target type.

[0009] Preferably, the terrain features include a combination of statistically derived static conditions of the terrain; The main vehicle trajectory shape includes the characteristics of the main vehicle's travel path; The target object trajectory shape includes the characteristics of the target vehicle's travel path; The conflict type characteristics include recording the changes in the relative positions of the subject vehicle and the target vehicle during the test condition; The target object type characteristics include pedestrians, bicycles and cars; The external input conditions include the measurement results of the changes in the horizontal and vertical relative distances and the horizontal and vertical relative speeds.

[0010] Preferably, the scenario files in the reconstruction of the representative simulation test scenario extracted based on the clustering results include OpenSceneRaio format files and SCE format files.

[0011] Compared with the prior art, the present invention has the following beneficial effects: The present invention significantly improves the quality of original data and reduces the impact of noise data by means of noise removal and trajectory smoothing, thereby improving the accuracy and reliability of subsequent analysis; And by using a comprehensive set of test scenario recognition rules, it can automatically identify high-value test scenarios and provide targeted test cases for the development of autonomous driving systems. These rules break away from the constraints of working condition definitions and lay the foundation for the mining of unknown unsafe working conditions. It ensures the diversity of test scenarios. Through similarity measurement and cluster analysis of test scenarios, it can effectively avoid duplication of test scenarios, ensure the diversity and comprehensiveness of test coverage, and provide a method for classifying and managing test scenarios, which helps to concentrate resources for in-depth research on the most challenging scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A system flow chart is built for the intelligent driving system test scenario of the present invention to collect data in the natural driving environment. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] See also Figure 1 The present invention provides a technical solution: a system for building intelligent driving system test scenarios based on data collected from a natural driving environment, the system sequentially including data preprocessing and quality control, automatic screening of test vehicles, automatic identification of high-value test scenarios based on the vehicle, similarity measurement of test scenarios, clustering based on test scenario similarity measurement, and reconstruction of representative simulation test scenarios based on extraction of clustering results.

[0015] The current mainstream methods for collecting data in natural driving environments include: road-side fixed equipment collection, which collects data through fixed cameras and sensors installed along the road; data collection vehicle collection, which uses vehicles equipped with a variety of sensors for mobile data collection; and drone collection, which uses drones for aerial photography to obtain a wider perspective and data.

[0016] Data preprocessing and quality control include trajectory noise point removal, noise target removal, trajectory smoothing and target merging.

[0017] Trajectory noise point deletion introduces the DBSCAN clustering algorithm to determine whether the trajectory is noise data based on the degree of clustering of trajectory points, delete the noise trajectory, and retain only one trajectory point when the vehicle is stationary to reduce redundancy; the DBSCAN clustering algorithm defines a cluster as the largest set of density-connected points, can divide areas with sufficiently high density into clusters, and can find clusters of any shape in the noisy spatial database.

[0018] Noise target deletion means that if the existence time of the target is less than 3 seconds, the target is considered as noise data and deleted; trajectory smoothing uses cubic spline interpolation method to deal with discontinuities or mutations in the collected data to obtain a smoother and more coherent vehicle trajectory; Data preprocessing and target merging for quality control include the following steps: Step 1, read the original data and the smoothed data; Step 2: extract the status data of each target object, including timestamp, speed, acceleration, position, heading angle and target type; Step 3: Analyze each target object and record the start and end states; Step 4: traverse all targets and predict the possible position range of the target within 5 seconds based on the target's end state. If the end state of another target coincides with this range, a merged map is constructed. Step five: process the data of the target object in the data according to the merging relationship.

[0019] There are a large number of traffic participants in the original collected data. Different traffic participants face different test scenarios. The first step in identifying high-value test scenarios is to screen the target objects that can be used as the main vehicle in the test. The test main vehicle is automatically screened according to the following three rules: Rule 1: The main vehicle must travel at least 100 meters; Rule 2: The main vehicle must continue to move for at least 15 seconds; Rule 3: The proportion of the main vehicle’s non-stationary time to the total time is at least 50%; If the above three rules are met at the same time, it can be used as the main vehicle; The automatic identification of high-value test scenarios based on the main vehicle is as follows: Based on each main vehicle selected above, seven screening rules are used to automatically identify high-value test scenarios. The seven screening rules include: Rule 1: Identification of the risk scenario of a forward collision with the main vehicle; Rule 2: Identify pedestrians or large vehicles in the front at a close distance; Rule 3: Intersection detection of the tracks of the host vehicle and the target vehicle; Rule 4: Identification of scenarios where traffic efficiency is limited; Rule 5: Identification of congested driving areas; Rule 6: Identification of driving area restricted scenes; Rule 7: Identify the rapid changes in the relative state of the target object; Any one or more of the above rules can be used as the basic test scenario; A large number of data fragments can be screened out in the above process. However, due to the long-tail effect of natural driving data, a large number of working conditions screened out have a high degree of similarity. In order to obtain diverse test scenarios, a similarity measurement method based on test scenario attributes is proposed. The similarity measurement of test scenario attributes is to measure the similarity of the basic test scenario attributes obtained above through six scenario attributes, where the six scenario attributes include: terrain features, main vehicle trajectory shape, target object trajectory shape, conflict type features, target object type features and external input conditions; Terrain characteristics include a combination of statistically derived static conditions of the terrain; The main vehicle trajectory shape includes the characteristics of the main vehicle's travel path; The target object trajectory shape includes the characteristics of the target vehicle's driving path; The conflict type characteristics include recording the changes in the relative positions of the subject vehicle and the target vehicle during the test condition; Object type features include pedestrians, bicycles, and cars; The external input conditions include the measurement results of the changes in the horizontal and vertical relative distances and the horizontal and vertical relative speeds.

[0020] Clustering based on test scene similarity measurement, that is, based on the above six scene attributes, the similarity between two targets can be calculated to obtain a similarity matrix. Combined with the neighbor propagation clustering analysis algorithm, cluster analysis can be performed on the test scenes. By calculating the similarity between scenes, scene groups with common characteristics can be identified, and the cluster center scenes can be extracted as high-value test scenes; Reconstruct representative simulation test scenarios based on clustering results, obtain representative high-value test scenarios based on clustering based on test scenario similarity metrics, extract their original data, and automatically convert the trajectory and speed data of all traffic participants into scenario files that can be used directly in the simulation environment, reproducing the movement trajectory and behavior of traffic participants in the simulation environment.

[0021] The generated simulation scene files support OpenScenraio format files and sce format files supported by SCANeR simulation software, which facilitates communication and application between different platforms.

[0022] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A system for building intelligent driving system test scenarios based on data collected from natural driving environments, characterized in that: The system includes data preprocessing and quality control, automatic screening of test vehicles, automatic identification of high-value test scenarios based on the vehicle, similarity measurement of test scenarios, clustering based on similarity measurement of test scenarios, and reconstruction of representative simulation test scenarios based on clustering results. The data preprocessing and quality control include trajectory noise point deletion, noise target deletion, trajectory smoothing and target merging; the trajectory noise point deletion introduces the DBSCAN clustering algorithm to determine whether the trajectory is noise data according to the degree of clustering of trajectory points, delete the noise trajectory, and only retain one trajectory point in the vehicle stationary stage to reduce redundancy; The noise target deletion means that if the existence time of the target is less than 3 seconds, the target is considered as noise data and deleted; the trajectory smoothing adopts the cubic spline interpolation method to deal with the discontinuity or mutation in the collected data to obtain a smoother and more coherent vehicle trajectory; The test vehicle is automatically screened according to the following three rules: Rule 1: The main vehicle must travel at least 100 meters; Rule 2: The main vehicle must continue to move for at least 15 seconds; Rule 3: The proportion of the main vehicle’s non-stationary time to the total time is at least 50%; If the above three rules are met at the same time, it can be used as the main vehicle; The automatic identification of high-value test scenarios based on the main vehicle is as follows: based on each of the above-screened main vehicles, seven screening rules are used to automatically identify high-value test scenarios. The seven screening rules include: Rule 1: Identification of the risk scenario of a forward collision with the main vehicle; Rule 2: Identify pedestrians or large vehicles in the front at a close distance; Rule 3: Intersection detection of the tracks of the host vehicle and the target vehicle; Rule 4: Identification of scenarios where traffic efficiency is limited; Rule 5: Identification of congested driving areas; Rule 6: Identification of driving area restricted scenes; Rule 7: Identify the rapid changes in the relative state of the target object; Any one or more of the above rules can be used as the basic test scenario; The similarity measurement of the test scene is to measure the similarity of the basic test scene attributes obtained above by six scene attributes, wherein the six scene attributes include: terrain characteristics, main vehicle trajectory shape, target object trajectory shape, conflict type characteristics, target object type characteristics and external input conditions; The clustering based on the test scene similarity measurement can calculate the similarity between two targets based on the above six scene attributes, thereby obtaining a similarity matrix. Combined with the neighbor propagation clustering analysis algorithm, the test scenes can be clustered and analyzed. By calculating the similarity between the scenes, scene groups with common characteristics can be identified, and the cluster center scenes can be extracted as high-value test scenes. The representative simulation test scenario reconstruction extracted based on clustering results is based on representative high-value test scenarios obtained by clustering based on test scenario similarity metrics, and its original data is extracted, and the trajectory and speed data of all traffic participants are automatically converted into scenario files that can be directly used in the simulation environment, reproducing the movement trajectory and behavior of traffic participants in the simulation environment.

2. The intelligent driving system test scenario building system based on natural driving environment data collection according to claim 1 is characterized by: The data preprocessing and target merging in quality control include the following steps: Step 1, read the original data and the smoothed data; Step 2: extracting the status data of each target object; Step 3: Analyze each target object and record the start and end states; Step 4: traverse all targets and predict the possible position range of the target within 5 seconds based on the target's end state. If the end state of another target coincides with this range, a merged map is constructed. Step 5: Process the data of the target object in the data according to the merging relationship.

3. The intelligent driving system test scenario building system based on natural driving environment data collection according to claim 2 is characterized by: The state data includes timestamp, speed, acceleration, position, heading angle and target type.

4. The intelligent driving system test scenario building system based on natural driving environment data collection according to claim 1 is characterized by: The terrain characteristics include a combination of statistically derived static conditions of the terrain; The main vehicle trajectory shape includes the characteristics of the main vehicle's travel path; The target object trajectory shape includes the characteristics of the target vehicle's travel path; The conflict type characteristics include recording the changes in the relative positions of the subject vehicle and the target vehicle during the test condition; The target object type characteristics include pedestrians, bicycles and cars; The external input conditions include the measurement results of the changes in the horizontal and vertical relative distances and the horizontal and vertical relative speeds.

5. The intelligent driving system test scenario building system based on natural driving environment data collection according to claim 1 is characterized by: The scenario files in the reconstruction of the representative simulation test scenario extracted based on the clustering results include OpenScenraio format files and sce format files.

Images