Vehicle-two-wheeled vehicle typical accident scene clustering method for intelligent vehicle test

Abstract

The invention relates to a vehicle-two-wheeled vehicle typical accident scene clustering method for intelligent vehicle testing, and the method comprises the following steps: S1, obtaining accident segments from historical accident data, and selecting a plurality of mixed elements; s2, respectively preprocessing the mixed elements to obtain a data set; s3, performing clustering processing based on the data set by adopting an adaptive competitive learning clustering algorithm to obtain a clustering result; s4, splitting one cluster into two sub-clusters according to a preset rule, performing clustering processing on the two sub-clusters and the unsplit cluster by adopting an adaptive competitive learning clustering algorithm, and updating a clustering result; s5, the step S4 is repeated, whether the adaptive competitive learning clustering algorithm converges is judged, and if the algorithm converges, a final clustering center and a sample label are output; otherwise, continuing to execute the steps S4-S5 until the adaptive competitive learning clustering algorithm converges; and S6, identifying a typical accident scene by analyzing the final clustering center and the sample label. The clustering result in the method is more comprehensive and objective.

Description

Technical Field

[0001] This manual relates to the field of intelligent vehicle scenario clustering, and in particular to a clustering method for typical accident scenarios of intelligent vehicle test vehicles - two-wheeled vehicles. Background Technology

[0002] Vehicle intelligence has become a key direction for the global automotive industry upgrade, but extreme weather, complex road conditions, and dynamically changing driving behaviors all pose serious challenges to the safety of autonomous driving systems. Scenarios are the core of autonomous driving development and testing. Among the various road traffic scenarios faced by autonomous vehicles, two-wheeled vehicles are a vulnerable group in road traffic due to their flexible movement trajectories, difficulty in predicting behavior, and weak protective measures. Interactions between vehicles and two-wheeled vehicles are often accompanied by a high risk of collision, making them a key edge and high-risk scenario that needs to be addressed in the current development and testing of autonomous driving systems. A comprehensive understanding of typical vehicle-two-wheeled vehicle interaction accident scenarios is helpful for the design and testing of autonomous driving perception and control algorithms.

[0003] Currently, the identification of typical test scenarios for autonomous driving mainly includes two methods: rule-based and clustering-based. While rule-based typical scenario identification methods are simple and easy to implement, they rely too heavily on expert experience to define rules for scenario extraction. Although highly targeted, they are also highly subjective, have rigid processes, and are difficult to automate in discovering unknown or potential typical patterns from data. In particular, when it is necessary to incorporate new scenario elements, the method often needs to be redesigned, resulting in poor scalability.

[0004] Therefore, identifying typical test scenarios for autonomous driving based on unsupervised clustering methods has become a mainstream trend. Scenario elements often exhibit significant "mixed attribute" characteristics, meaning they simultaneously contain continuous variables such as speed and distance, as well as discrete categorical variables such as ambient lighting. Real-world accident data suffers from severe sample imbalance due to the frequency of occurrence. Conventional clustering algorithms (such as K-means and K-prototypes) tend to identify high-frequency, common scenarios while ignoring less frequent but crucial high-risk edge scenarios.

[0005] Autonomous driving test scenarios include two main categories of elements: static and dynamic. Static elements include road conditions and environmental information, while dynamic elements include the motion and interaction behavior of the test vehicle and other two-wheeled vehicles. Due to the difficulty of data fusion and processing, current research mostly selects only a single type of continuous element for clustering, or simply converts discrete elements into numerical processing. This results in clustering results that cannot accurately represent the coupling relationships between heterogeneous elements in real traffic scenarios, making it difficult to recreate complex real-world accident patterns. Summary of the Invention

[0006] This manual provides a clustering method for typical accident scenarios of intelligent vehicle test vehicles - two-wheeled vehicles, which solves the problem in existing technologies that make it difficult to effectively integrate heterogeneous data, resulting in incomplete extraction of typical scenarios.

[0007] This manual adopts the following technical solution: a clustering method for typical accident scenarios of intelligent vehicle test vehicles and two-wheeled vehicles, including the following steps: S1: Obtain accident fragments from historical accident data, and select multiple mixed elements from the accident fragments; S2: Preprocess the multiple mixed elements separately to obtain the preprocessed dataset; S3: Based on the preprocessed dataset, an adaptive competitive learning clustering algorithm is used to perform clustering to obtain clustering results, which include multiple clusters; S4: According to the preset rules, split one of the clusters into two sub-clusters, and then use the adaptive competitive learning clustering algorithm to cluster the two sub-clusters and the unsplit cluster to obtain a new clustering result; S5: Repeat step S4 and determine whether the adaptive competitive learning algorithm has converged. If the adaptive competitive learning algorithm has converged, output the final cluster centers and sample labels. If the adaptive competitive learning algorithm fails to converge, continue executing steps S4-S5 until the adaptive competitive learning algorithm converges. S6: Identify typical accident scenarios by analyzing the final cluster centers and sample labels.

[0008] Based on the above technical means, scene fragments are screened from historical accident data, mixed elements are extracted from the scene fragments and preprocessed, and finally analyzed by an adaptive competitive learning clustering algorithm. The whole process is unsupervised learning, which can reduce the reliance on manual rules and prior knowledge and improve the objectivity and efficiency of scene clustering.

[0009] In this application, the mixed elements extracted from historical accident data are clustered using an adaptive competitive learning algorithm, which can effectively identify small sample long-tail data and solve the problem of loss of rare high-risk scenarios caused by severe imbalance in accident data.

[0010] Furthermore, S3 includes the following steps: S31: Set the initial number of seed points and learning rate parameters, and randomly select samples from the preprocessed dataset as initial cluster centers; S32: Input a certain data point from the preprocessed dataset into the adaptive competitive learning algorithm, obtain the distance between the certain data point and various current sub-points, and record the seed point with the smallest distance as the winning point, and the other seed points as competitive points; S33: Adjust and update the winning point position so that the winning point is closer to a certain data; Adjust and update the position of the corresponding competitive point so that the corresponding competitive point is far away from a certain data.

[0011] Based on the aforementioned technical means, by setting the initial number of seed points and learning rate parameters, initial cluster centers are randomly selected from the preprocessed dataset. Winning points and competing points are dynamically determined based on the distance between data points and seed points. In turn, the winning points are adaptively adjusted to move closer to the data points and the competing points are moved further away from the data points. This achieves refined iterative optimization of the cluster centers, improves the adaptability and convergence efficiency of the clustering process, alleviates the problem of losing typical rare scenarios due to imbalanced accident data, and enhances the ability of clustering results to represent real and complex accident patterns.

[0012] Furthermore, the S32 formula is as follows: , in, This indicates that the i-th seed point corresponds to a certain data position. The winning symbol This represents the position of the i-th seed point. This refers to a specific data position. When the distance between the seed point and this specific data position is minimized, the seed point at this point is the winning point. It is 1; otherwise, it is a competitive point. 0.

[0013] Furthermore, the adjustment in S33 is specifically as follows: , in, K is the learning rate, and K is the number of seed points. Let η be the competition point penalty coefficient, and η be a coefficient related to the competition point penalty coefficient. If it is 1, then This indicates the winning position of a certain data point. Indicates the adjusted winning position; if If it is not 1, then This indicates the location of a contention point for a given piece of data. This indicates the adjusted position of the competitive point.

[0014] Furthermore, the specific details of the competition point penalty coefficient are as follows: , in, Indicates the winning point position. Indicates the location of the competing point.

[0015] Furthermore, S4 includes the following steps: S41: Count the number of times each type of point becomes a winning point; S42: Determine the cluster of the seed point that has become the winning point the most times, and count the density gap of the cluster; S43: Split the cluster containing the seed point with the most winning points into two sub-clusters and generate two new seed points, which are located in the two new sub-clusters respectively, to replace the original seed point with the most winning points, and update the seed point number K value.

[0016] Based on the aforementioned technical means, by statistically analyzing the number of wins of each cluster center (seed point) and selecting the cluster with the most wins for splitting, it is possible to adaptively identify and process cluster structures with uneven data distribution. Specifically, for high-frequency accident scenario clusters with a large sample size and potential heterogeneous sub-patterns, this mechanism can proactively split them into two more homogeneous sub-clusters based on the density gaps within the cluster, while generating new cluster centers. This avoids the sparse but crucial scenario patterns being masked by the dominant pattern due to uneven data distribution caused by improper initial cluster center settings or a fixed number of clusters.

[0017] Furthermore, the formula for generating new seed points in S43 is as follows: , in, This represents the set of points from which new seed points are generated. This represents the number of wins for the j-th seed. This represents the density gap of the cluster containing the j-th seed point.

[0018] Furthermore, the mixed elements in S2 include discrete data and continuous data.

[0019] Based on the aforementioned technical means, the mixed elements, including discrete and continuous data, can ensure the integrity and authenticity of the scene description.

[0020] Furthermore, the preprocessing in S2 includes: For continuous data, the maximum-minimum normalization method is used to map it to the [0,1] interval; For discrete data, one-hot encoding is used to convert it into a numerical vector.

[0021] Based on the above technical means, continuous data is mapped to the [0, 1] interval through the maximum and minimum value normalization method to eliminate the influence of dimensions; discrete data is converted into 0 / 1 vectors through one-hot encoding to ensure the equivalence of distance calculation between variables.

[0022] Furthermore, S5 also includes: after the algorithm converges, using the silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin index to evaluate and verify the output clustering results.

[0023] Based on the above technical means, the silhouette coefficient, Calinski-Harabasz index and Davies-Bouldin index are comprehensively used to evaluate the clustering results, objectively quantifying the intra-cluster compactness and inter-cluster separation of the clustering results, and can more robustly and reliably determine the optimal number of clusters.

[0024] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: 1. Scene fragments are selected from historical accident data, and mixed elements are extracted and preprocessed from the scene fragments. Finally, the analysis is performed through an adaptive competitive learning clustering algorithm. The whole process is unsupervised learning, which can reduce the reliance on manual rules and prior knowledge and improve the objectivity and efficiency of scene clustering.

[0025] 2. In this application, the mixed elements extracted from historical accident data are clustered by an adaptive competitive learning algorithm, which can effectively identify small sample long-tail data and solve the problem of loss of rare high-risk scenes caused by severe imbalance in accident data. Attached Figure Description

[0026] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of this embodiment; Figure 2 This is a schematic diagram of the clustering process in this embodiment; Figure 3 This is a schematic diagram of the cluster splitting process in this embodiment; Figure 4 This is a graph showing the variation of the silhouette coefficient with the number of clusters in this embodiment; Figure 5 This is a graph showing the variation of the CH index with the number of clusters in this embodiment; Figure 6 This is a graph showing how the DB index changes with the number of clusters in this embodiment. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.

[0028] The embodiments of the present invention will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0029] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0030] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.

[0031] In the embodiments of this application, unless otherwise explicitly specified and limited, the term "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral part; it can be a direct connection or an indirect connection through an intermediate medium.

[0032] In embodiments of this application, 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 limitation, 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 that element.

[0033] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0034] Example 1 like Figure 1 As shown, this implementation provides a clustering method for typical accident scenarios of intelligent vehicle test vehicles - two-wheeled vehicles, including the following steps: S1: Obtain accident fragments from historical accident data, and select multiple mixed elements from the accident fragments.

[0035] In this preferred embodiment, S1 specifically involves: firstly, selecting and extracting typical vehicle-to-two-wheeled vehicle accident scene fragments from the Chinese road natural driving dataset, using the following extraction rules: Extract data segments from the accident dataset where the target object type is identified as "two-wheeled vehicle"; Identify the behavioral intentions and relative movement directions of the main vehicle and the two-wheeled vehicle, and verify unreasonable data in the dataset; The hybrid elements that can fully describe the vehicle-two-wheeler interaction scenario were extracted from the data fragments, as shown in Table 1.

[0036] S2: Preprocess the multiple mixed elements separately to obtain the preprocessed dataset.

[0037] In this preferred embodiment, continuous variables such as vehicle speed are mapped to the [0, 1] interval through Min-Max normalization to eliminate the influence of dimensions; discrete variables such as the behavior of the main vehicle and time are converted into 0 / 1 vectors through one-hot encoding to ensure the equivalence of distance calculation between variables; and the preprocessed mixed elements are sorted to obtain the preprocessed dataset.

[0038] S3: Based on the preprocessed dataset, an adaptive competitive learning clustering algorithm is used to perform clustering to obtain clustering results, which include multiple clusters.

[0039] like Figure 2 As shown, in this preferred embodiment, S3 includes the following steps: S31: Set the initial seed point number K and learning rate parameters, and randomly select samples from the preprocessed dataset as initial cluster centers; S32: Input a certain data point from the preprocessed dataset into the adaptive competitive learning algorithm, obtain the distance between the certain data point and various current sub-points, and record the seed point with the smallest distance as the winning point, and the other seed points as competitive points; In this preferred embodiment, the determination and marking of the competition point and the winning point in S32 are specifically as follows: , in, This indicates that the i-th seed point corresponds to a certain data position. The winning symbol This represents the position of the i-th seed point. This refers to a specific data position. When the distance between the seed point and this specific data position is minimized, the seed point at this point is the winning point. It is 1; otherwise, it is a competitive point. 0.

[0040] S33: Adjust and update the winning point position so that the winning point is closer to a certain data; Adjust and update the position of the corresponding competitive point so that the corresponding competitive point is far away from a certain data.

[0041] In this preferred embodiment, the adjustment method and adjustment amount of S33 are as follows: , in, K is the learning rate, and K is the number of seed points. Let η be the competition point penalty coefficient, and η be a coefficient related to the competition point penalty coefficient. If it is 1, then This indicates the winning position of a certain data point. Indicates the adjusted winning position; if If it is not 1, then This indicates the location of a contention point for a given piece of data. This indicates the adjusted position of the competitive point.

[0042] The method for obtaining the competition point penalty coefficient is as follows: , in, Indicates the winning point position. Indicates the location of the competing point.

[0043] In this preferred embodiment, the adaptive competitive learning algorithm can adaptively increase the number of seed points until a suitable size is reached, and the learning rate... α c Set it to 0.001. Set to 0.001 to ensure the algorithm does not terminate prematurely; set the frequency threshold to θ = 0.02. N This means that if a seed point wins less than 1 / 50 of the total number of data points, it will be removed.

[0044] S4: According to the preset rules, split one of the clusters into two sub-clusters, and then use the adaptive competitive learning clustering algorithm to cluster the two sub-clusters and the unsplit clusters, and update the clustering results.

[0045] like Figure 4As shown, in this preferred embodiment, to achieve clustering of imbalanced data, multi-prototype clustering is adopted, where each cluster is represented by one or more sub-clusters, and the clusters to be split into multiple sub-clusters are selected by the following method: S41: Count the number of times each sub-point becomes a winning point; S42: Determine the cluster of the seed point that has become the winning point the most times, and count the density gap of the cluster; S43: Split the cluster containing the seed point with the most winning points into two sub-clusters and generate two new seed points, which are located in the two new sub-clusters respectively, to replace the original seed point with the most winning points, and update the seed point number K value.

[0046] In this embodiment, by counting the number of wins of each cluster center (seed point) and selecting the cluster with the most wins for splitting, it can adaptively identify and process cluster structures with uneven data distribution. Specifically, for high-frequency accident scene clusters with huge sample sizes and possible heterogeneous sub-patterns, this mechanism can actively split them into two more homogeneous sub-clusters based on the density gaps within the cluster, while generating new cluster centers. This avoids the sparse but critical scene patterns being masked by the dominant pattern due to uneven data distribution caused by improper initial cluster center settings or a fixed number of clusters.

[0047] S5: Repeat step S4 and determine whether the adaptive competitive learning algorithm has converged. If the adaptive competitive learning algorithm has converged, output the final cluster centers and sample labels. If the adaptive competitive learning algorithm fails to converge, continue executing steps S4-S5 until the adaptive competitive learning algorithm converges. In this preferred embodiment, the convergence threshold of the adaptive competitive learning algorithm is set to... = 0.01, number of iterations E It is 1000.

[0048] In this preferred embodiment, in order to determine the optimal number of clusters K, the silhouette coefficient, Calinski-Harabasz (CH) index, and Davies-Bouldin (DB) index are used for evaluation.

[0049] Among them, the silhouette coefficient is an indicator used to evaluate the quality of clustering results, measuring the compactness and separation of samples in the cluster; the CH coefficient is an indicator used to evaluate the quality of clustering results, based on the ratio of intra-cluster dispersion to inter-cluster dispersion, and the larger the value, the better the clustering effect; the core idea of ​​the DB index is to calculate the similarity between each cluster and its most similar cluster, and then measure the quality of the entire clustering result by averaging all similarities. If the similarity between clusters is higher (the DB index is higher), it means that the distance between clusters is smaller (intuitively, only things that are closer together are similar), then the clustering result is worse, and vice versa.

[0050] like Figures 4-6 As shown, in this embodiment, the contour coefficient reaches its peak at K=7; the CH index reaches its peak at K=2; and the DB index reaches its lowest point at K=7.

[0051] Taking all the above indicators into consideration, this invention ultimately selects 7 categories as the optimal number of clusters.

[0052] The characteristic parameters of the seven typical vehicle-two-wheeler accident scenarios obtained are shown in Table 2.

[0053]

[0054] S6: Identify typical accident scenarios by analyzing the final cluster centers and sample labels.

[0055] In summary, this embodiment effectively solves the problems of clustering difficulties and incomplete extraction of typical features caused by heterogeneous data attributes in vehicle-to-two-wheeled vehicle interaction scenarios. It is applied to the clustering of typical vehicle-to-two-wheeled vehicle accident scenarios, enriches the existing test scenario library for vulnerable traffic participants in intelligent connected vehicles, and has broad application prospects in the development and safety testing of autonomous driving functions.

[0056] The typical scenario library constructed by this invention enriches the test cases for autonomous vehicles targeting vulnerable road users, and can provide targeted test inputs for the game interaction capabilities of autonomous driving control algorithms, which has important application value in improving the safety of autonomous driving.

[0057] The above embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention.

Claims

1. A clustering method for typical accident scenarios of intelligent vehicle test vehicles - two-wheeled vehicles, characterized in that, Includes the following steps: S1: Obtain accident fragments from historical accident data, and select multiple mixed elements from the accident fragments; S2: Preprocess the multiple mixed elements separately to obtain the preprocessed dataset; S3: Based on the preprocessed dataset, an adaptive competitive learning clustering algorithm is used to perform clustering to obtain clustering results, which include multiple clusters; S4: According to the preset rules, split one of the clusters into two sub-clusters, and then use the adaptive competitive learning clustering algorithm to cluster the two sub-clusters and the unsplit clusters, and update the clustering results; S5: Repeat step S4 and determine whether the adaptive competitive learning algorithm has converged. If the adaptive competitive learning algorithm has converged, output the final cluster centers and sample labels. If the adaptive competitive learning algorithm fails to converge, continue executing steps S4-S5 until the adaptive competitive learning algorithm converges. S6: Identify typical accident scenarios by analyzing the final cluster centers and sample labels.

2. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 1, characterized in that, S3 includes the following steps: S31: Set the initial number of seed points and learning rate parameters, and randomly select samples from the preprocessed dataset as initial cluster centers; S32: Input a certain data point from the preprocessed dataset into the adaptive competitive learning algorithm, obtain the distance between the certain data point and various current sub-points, and record the seed point with the smallest distance as the winning point, and the other seed points as competitive points; S33: Adjust and update the winning point position so that the winning point is closer to a certain data; Adjust and update the position of the corresponding competitive point so that the corresponding competitive point is far away from a certain data.

3. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 2, characterized in that, The S32 formula is as follows: ， in, This indicates that the i-th seed point corresponds to a certain data position. The winning symbol This represents the position of the i-th seed point. This refers to a specific data position. When the distance between the seed point and this specific data position is minimized, the seed point at this point is the winning point. It is 1; otherwise, it is a competitive point.

0.

4. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 3, characterized in that, The specific adjustment in S33 is as follows: ， in, K is the learning rate, and K is the number of seed points. Let η be the competition point penalty coefficient, and η be a coefficient related to the competition point penalty coefficient. If it is 1, then This indicates the winning position of a certain data point. Indicates the adjusted winning position; if If it is not 1, then This indicates the location of a contention point for a given piece of data. This indicates the adjusted position of the competitive point.

5. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 4, characterized in that, The specific details of the competition point penalty coefficient are as follows: ， in, Indicates the winning point position. Indicates the location of the competing point.

6. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 2, characterized in that, S4 includes the following steps: S41: Count the number of times each sub-point becomes a winning point; S42: Determine the cluster of the seed point that has become the winning point the most times, and count the density gap of the cluster; S43: Split the cluster containing the seed point with the most winning points into two sub-clusters and generate two new seed points, which are located in the two new sub-clusters respectively, to replace the original seed point with the most winning points, and update the seed point number K value.

7. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 6, characterized in that, The formula for generating new seed points in S43 is as follows: ， in, This represents the set of points from which new seed points are generated. This represents the number of wins for the j-th seed. This represents the density gap of the cluster containing the j-th seed point.

8. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 1, characterized in that, The mixed elements in S2 include discrete data and continuous data.

9. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 8, characterized in that, The preprocessing in S2 includes: For continuous data, the maximum-minimum normalization method is used to map it to the [0,1] interval; For discrete data, one-hot encoding is used to convert it into a numerical vector.

10. The clustering method for typical accident scenarios of intelligent vehicle test vehicles-two-wheeled vehicles according to claim 1, characterized in that, S5 further includes: after the algorithm converges, using the silhouette coefficient, Calinski-Harabasz index, and Davies-Bouldin index to evaluate and verify the output clustering results.

Images