An automatic driving vehicle high-risk scene identification system and method

Abstract

This invention discloses a system and method for identifying high-risk accident scenarios for autonomous vehicles, relating to the field of autonomous vehicle safety assessment technology. It includes: a data acquisition and preprocessing module for acquiring accident data and determining scenario factors; a risk factor identification module for establishing a machine learning-based risk level prediction model and identifying key risk factors; an association rule mining module for mining coupling patterns between factors in high-risk accidents and generating strong association rules; and a scenario profile generation module for constructing high-risk scenario profiles based on strong association rules to obtain the identification results of high-risk accident scenarios for autonomous vehicles. This invention can provide support for the construction of test scenarios and the optimization of safety algorithms for autonomous vehicles.

Description

Technical Field

[0001] This invention belongs to the field of autonomous vehicle safety assessment technology, and in particular relates to a system and method for identifying high-risk accident scenarios of autonomous vehicles. Background Technology

[0002] The development of autonomous vehicles has offered new possibilities for improving road traffic safety. However, autonomous driving systems need to cope with complex operating conditions such as variable weather, complex traffic flow, and diverse driving tasks, and traditional vehicle testing methods can no longer meet their safety verification requirements. Conducting testing and evaluation of autonomous vehicles based on high-risk accident scenarios has become an important means of improving their safety. Currently, autonomous vehicle testing scenarios mainly originate from standards and regulations, natural driving data, traffic accident data, and simulation data. Among these, traffic accident data has a higher evaluation value because it can realistically reflect dangerous operating conditions. Existing research mostly focuses on constructing high-risk scenarios for traditional vehicles, using accident data to extract typical collision scenarios through clustering methods, such as car-pedestrian and car-two-wheeler accidents. The few studies on autonomous vehicles also mainly use clustering algorithms, relying heavily on subjective judgment in the selection of scenario factors and rarely considering the technical characteristic variables unique to autonomous driving (such as sensor type, driving mode, vehicle model, and year). Furthermore, clustering methods can only discover similar sample groups, making it difficult to intuitively present risk patterns coupled with multiple factors and failing to deeply reveal the complex mechanisms of accident occurrence.

[0003] Although existing research has attempted to apply association rules to road traffic accident analysis, the following technical challenges remain in the field of identifying high-risk accident scenarios for autonomous vehicles: First, the scenario factor system is incomplete, lacking a systematic inclusion of autonomous driving technology characteristic variables, resulting in insufficient explanatory power for accident risks; second, accident data suffers from class imbalance, limiting the ability of traditional models to identify high-risk accidents; third, existing methods mostly focus on single-factor influence analysis, failing to fully explore higher-order coupling relationships between multiple factors, making it difficult to construct interpretable high-risk scenario profiles. These problems restrict the accurate construction of autonomous vehicle test scenarios and the targeted optimization of safety algorithms, necessitating a high-risk scenario identification method that can comprehensively consider multi-dimensional factors, handle data imbalance, and reveal the coupling effects between factors. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a system and method for identifying high-risk accident scenarios for autonomous vehicles. Specifically, the system for identifying high-risk accident scenarios for autonomous vehicles includes: The data acquisition and preprocessing module is used to acquire accident data from autonomous vehicles and determine scenario factors; The risk factor identification module, connected to the data acquisition and preprocessing module, is used to receive the accident data and the scenario factors, establish a machine learning-based prediction model for the degree of accident risk of autonomous vehicles, and identify key risk factors. The association rule mining module is connected to the risk factor identification module and is used to receive the key risk factors, mine the coupling patterns between factors in high-risk accidents through association rule mining, and generate strong association rules. The scene profile generation module is connected to the association rule mining module. It is used to receive the strong association rules, construct a high-risk scene profile of autonomous vehicle accidents based on the coupling mode, and obtain the identification result of high-risk scene of autonomous vehicle accidents.

[0005] Preferably, the data acquisition and preprocessing module includes: The data cleaning unit is used to collect collision accident reports from autonomous vehicles and clean incomplete data. The feature expansion unit, connected to the data cleaning unit, is used to expand the technical feature variables of autonomous vehicles based on traditional scenario factors; the technical feature variables of autonomous vehicles include vehicle model year, sensor type and driving mode; The risk level classification unit, connected to the feature extension unit, is used to classify the degree of accident risk into three levels: high risk, medium risk, and low risk.

[0006] Preferably, the risk factor identification module includes: The data balancing unit is used to balance imbalanced accident data using a random oversampling instance method to generate balanced accident data. The model building unit, connected to the data balancing unit, is used to build an accident risk prediction model based on LightGBM based on the balanced accident data. The key factor identification unit, connected to the model building unit, is used to calculate the contribution value of each scenario factor to the output of the prediction model by combining the SHAP algorithm, and to select key risk factors according to the magnitude of the contribution value.

[0007] Preferably, the association rule mining module includes: The rule mining unit is used to mine association rules for key risk factors in high-risk accidents based on the Apriori algorithm. The rule filtering unit, connected to the rule mining unit, is used to filter strongly correlated rules that meet the preset minimum support, minimum confidence, and lift greater than 1. The network construction unit, connected to the rule filtering unit, is used to construct an accident key risk factor network with key risk factors as nodes and lift degree as edge weights based on the filtered strong correlation rules.

[0008] Preferably, the scene profile generation module includes: The feature combination extraction unit is used to extract the core risk feature combination and synergistic factors among the factors based on the rules with the highest lift in the strong association rules. The scenario summarization unit, connected to the feature combination extraction unit, is used to summarize and form a typical high-risk accident scenario profile based on the core risk feature combination and synergistic factors.

[0009] This invention also provides a method for identifying high-risk accident scenarios for autonomous vehicles, comprising: Acquire accident data from autonomous vehicles and determine scenario factors; Based on the accident data and scenario factors, a machine learning-based model for predicting the degree of accident risk of autonomous vehicles is established, and key risk factors are identified. Based on the key risk factors, strong association rules are generated by mining the coupling patterns between factors in high-risk accidents through association rule mining. Based on the strong association rules, a high-risk scenario profile for autonomous vehicle accidents is constructed, and the identification results of high-risk scenarios for autonomous vehicle accidents are obtained.

[0010] Preferably, the process of acquiring accident data from autonomous vehicles and determining scenario factors includes: Collect collision accident reports from autonomous vehicles and clean incomplete data; In addition to traditional scenario factors, the technical characteristic variables of autonomous vehicles are expanded; among them, the technical characteristic variables of autonomous vehicles include vehicle model year, sensor type and driving mode; The degree of accident risk is divided into three levels: high risk, medium risk, and low risk.

[0011] Preferably, the process of establishing a machine learning-based model for predicting the degree of accident risk of autonomous vehicles and identifying key risk factors includes: The imbalanced accident data is balanced using a random oversampling instance method to generate balanced accident data. Based on the balanced accident data, an accident risk prediction model based on LightGBM is constructed. The contribution value of each scenario factor to the prediction model output is calculated by combining the SHAP algorithm, and key risk factors are selected based on the magnitude of the contribution value.

[0012] Preferably, the process of generating strong association rules by mining the coupling patterns between factors in high-risk accidents includes: Based on the Apriori algorithm, association rules are mined for key risk factors in high-risk accidents. Filter out strong association rules that meet the preset minimum support, minimum confidence, and lift greater than 1.

[0013] Preferably, after generating the strong association rule, the method further includes: Based on the selected strong association rules, an accident key risk factor network is constructed with the key risk factors as nodes and the lift degree as the edge weight.

[0014] Compared with the prior art, the present invention has the following advantages and technical effects: This invention improves the scenario factor system and enhances the accuracy of accident risk characterization by introducing technical characteristic variables unique to autonomous vehicles. It employs the Random Over-Sampling Examples (ROSE) method to address data imbalance, combining the LightGBM model and the SHAP interpretable algorithm to accurately identify key risk factors while revealing the direction and extent of each factor's impact on accident risk. Furthermore, it utilizes association rules to mine coupling patterns among factors in high-risk accidents and constructs a network of key accident risk factors, intuitively presenting the synergistic effects of multiple factors, ultimately generating four typical high-risk scenario profiles. Compared to existing clustering methods, this invention not only discovers strong correlations between factors but also quantifies the degree to which each combination contributes to high-risk accidents, solving the problem of existing technologies' inability to explain multi-factor coupling mechanisms. This provides a more targeted basis for constructing test scenarios for autonomous vehicles, optimizing safety algorithms, and assessing risks. Attached Figure Description

[0015] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of the interpretable machine learning and association rule combination model framework according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the contribution of the SHAP factor in an embodiment of the present invention. Figure 3 This is a schematic diagram of the density distribution of valuable strong association rules in an embodiment of the present invention; Figure 4 This is a schematic diagram of the network of key risk factors for high-risk accidents in an embodiment of the present invention; Figure 5 This is a schematic diagram of a high-risk scene in an embodiment of the present invention. Detailed Implementation

[0016] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0018] Example 1 This embodiment provides a high-risk accident scene identification system for autonomous vehicles, including: The data acquisition and preprocessing module is used to acquire accident data from autonomous vehicles and determine scenario factors; The risk factor identification module, connected to the data acquisition and preprocessing module, is used to receive accident data and scenario factors, establish a machine learning-based prediction model for the degree of accident risk of autonomous vehicles, and identify key risk factors. The association rule mining module, connected to the risk factor identification module, is used to receive key risk factors and mine the coupling patterns between factors in high-risk accidents through association rule mining to generate strong association rules. The scene profile generation module is connected to the association rule mining module. It is used to receive strong association rules, construct high-risk scene profiles of autonomous vehicle accidents based on the coupling mode, and obtain the identification results of high-risk scene profiles of autonomous vehicle accidents.

[0019] Furthermore, the data acquisition and preprocessing module includes: The data cleaning unit is used to collect collision accident reports from autonomous vehicles and clean incomplete data. The feature expansion unit, connected to the data cleaning unit, is used to expand the technical feature variables of autonomous vehicles based on traditional scenario factors. The technical feature variables of autonomous vehicles include vehicle model year, sensor type, and driving mode. The risk level classification unit, connected to the feature extension unit, is used to classify the degree of accident risk into three levels: high risk, medium risk, and low risk.

[0020] Furthermore, this embodiment expands upon traditional scenario factors by adding technical characteristic variables of autonomous vehicles, including vehicle model year, sensor type, and driving mode, through the data acquisition and preprocessing module and its feature extension unit. This improves the scenario factor system, enabling accident risk analysis to fully reflect the technical characteristics of the autonomous driving system and enhancing the accuracy and explanatory power of accident risk characterization.

[0021] Furthermore, the risk factor identification module includes: The data balancing unit is used to balance imbalanced accident data using a random oversampling instance method to generate balanced accident data. The model building unit, connected to the data balancing unit, is used to build an accident risk prediction model based on LightGBM based on the balanced accident data. The key factor identification unit, connected to the model building unit, is used to calculate the contribution value of each scenario factor to the output of the prediction model by combining the SHAP algorithm, and select key risk factors according to the magnitude of the contribution value.

[0022] Furthermore, this embodiment uses the ROSE method to balance imbalanced accident data through the data balancing unit in the risk factor identification module, which solves the model bias problem caused by the small proportion of high-risk accident samples and improves the model's sensitivity and generalization ability in identifying high-risk accidents.

[0023] Furthermore, this embodiment constructs an accident risk prediction model based on LightGBM through the model building unit and key factor identification unit in the risk factor identification module, and calculates the contribution value of each scenario factor to the model output by combining the SHAP algorithm. This not only realizes the rapid classification and prediction of high-dimensional data, but also quantifies the direction and degree of influence of each factor on accident risk, enhances the interpretability of the model, and provides accurate key factor input for subsequent association rule mining.

[0024] Furthermore, the association rule mining module includes: The rule mining unit is used to mine association rules for key risk factors in high-risk accidents based on the Apriori algorithm. The rule filtering unit, connected to the rule mining unit, is used to filter strongly correlated rules that meet the preset minimum support, minimum confidence, and lift greater than 1. The network construction unit, connected to the rule filtering unit, is used to construct an accident key risk factor network with key risk factors as nodes and lift degree as edge weights based on the filtered strong correlation rules.

[0025] Furthermore, this embodiment uses the rule mining unit and rule filtering unit in the association rule mining module to mine association rules between key risk factors in high-risk accidents based on the Apriori algorithm, and filters strong association rules that meet the requirements of support, confidence and lift greater than 1. This breaks through the limitation of traditional cluster analysis that can only find similar sample groups, intuitively presents the high-order coupling mode between multiple factors, and quantifies the improvement effect of each factor combination on high-risk accidents.

[0026] Furthermore, this embodiment constructs a network of key risk factors for accidents using the network construction unit in the association rule mining module, with key risk factors as nodes and lift degree as edge weights. It visualizes the association strength and interaction between factors, providing an intuitive topological structure to support the understanding of the complex mechanisms of accident occurrence.

[0027] Furthermore, the scene profile generation module includes: The feature combination extraction unit is used to extract the core risk feature combinations and synergistic factors among factors based on the rules with the highest lift in the strong association rules. The scenario summarization unit, connected to the feature combination extraction unit, is used to summarize typical high-risk accident scenario profiles based on the combination of core risk features and synergistic factors.

[0028] Furthermore, this embodiment extracts core risk feature combinations and synergistic factors based on strong association rules ranked high in lift degree through the feature combination extraction unit and scene summarization unit in the scene profile generation module. It then summarizes these into typical high-risk accident scene profiles, transforming abstract association rules into scene descriptions that can be directly applied to autonomous vehicle testing. This provides a concrete basis for test scene construction and safety algorithm optimization.

[0029] In summary, the system in this embodiment, through the sequential connection and collaborative work of various modules in the overall system architecture, forms a complete high-risk scenario identification process from data collection, risk factor identification, association rule mining to scenario profile generation. It solves the technical problems of incomplete factor system, data imbalance, and difficulty in interpreting multi-factor coupling mechanism in the prior art, and realizes accurate identification and interpretable output of high-risk scenarios of autonomous vehicle accidents.

[0030] Example 2 Based on the same inventive concept, this embodiment also provides a method for identifying high-risk accident scenarios for autonomous vehicles, including: Acquire accident data from autonomous vehicles and determine scenario factors; Based on accident data and scenario factors, a machine learning-based model for predicting the degree of accident risk of autonomous vehicles is established, and key risk factors are identified. Based on key risk factors, strong association rules are generated by mining the coupling patterns between factors in high-risk accidents through association rule mining. Based on strong association rules, a high-risk scenario profile for autonomous vehicle accidents is constructed, and the identification results of high-risk scenarios for autonomous vehicle accidents are obtained.

[0031] Furthermore, the process of acquiring accident data from autonomous vehicles and identifying scenario factors includes: Collect collision accident reports from autonomous vehicles and clean incomplete data; In addition to traditional scenario factors, the technical characteristic variables of autonomous vehicles are expanded; among them, the technical characteristic variables of autonomous vehicles include vehicle model year, sensor type and driving mode; The degree of accident risk is divided into three levels: high risk, medium risk, and low risk.

[0032] Furthermore, the process of establishing a machine learning-based model for predicting the degree of accident risk for autonomous vehicles and identifying key risk factors includes: The imbalanced accident data is balanced using a random oversampling instance method to generate balanced accident data. Based on the balanced accident data, an accident risk prediction model based on LightGBM is constructed. The contribution of each scenario factor to the prediction model output is calculated using the SHAP algorithm, and key risk factors are selected based on the magnitude of the contribution value.

[0033] Furthermore, the process of generating strong association rules by mining the coupling patterns between factors in high-risk accidents includes: Based on the Apriori algorithm, association rules are mined for key risk factors in high-risk accidents. Filter out strong association rules that meet the preset minimum support, minimum confidence, and lift greater than 1.

[0034] Furthermore, after generating strong association rules, the following are also included: Based on the selected strong correlation rules, an accident key risk factor network is constructed with key risk factors as nodes and lift degree as edge weight.

[0035] The method for identifying high-risk accident scenarios of autonomous vehicles provided in this embodiment has all the advantages of the autonomous vehicle accident high-risk accident scenario identification system provided in Embodiment 1.

[0036] As a preferred implementation method, the high-risk scenario identification method for autonomous vehicle accidents provided in this embodiment uses data from autonomous vehicle collision accident reports published by the California Department of Motor Vehicles from January 2018 to December 2024. After removing incomplete accident data, a total of 702 accident data involving autonomous vehicles were used for modeling. Regarding vehicle perception mechanisms, sensor type was added as an extended variable; regarding road information, fields such as "one-way / two-way," "physical separation," and "number of lanes" were added. Ultimately, a total of 17 scenario factors were used as independent variables, divided into four categories: vehicle dimension, collision detail dimension, environment dimension, and road dimension.

[0037] To distinguish high-risk scenarios from numerous accident scenarios, the degree of accident risk was measured using personal injury and vehicle damage as joint dependent variables. Accidents were categorized into three types: high-risk (1 or more injuries or severe vehicle damage), medium-risk (0 injuries and minor vehicle damage), and low-risk (0 injuries and no vehicle damage), accounting for 26.50%, 67.81%, and 5.69% of the total number of accidents, respectively. Table 1 provides a statistical description of 17 scenario factors. Driving mode refers to whether the autonomous vehicle is in autonomous driving mode; "autonomous" means the vehicle is in normal operation of the autonomous driving system; "manual" means the vehicle has exited autonomous driving mode and is being manually controlled; "autonomous vehicle" refers to the autonomous vehicle involved in the accident, and vehicles colliding with the autonomous vehicle are defined as the collision object.

[0038] Table 1 It should be noted that continuous variables need to be discretized and binned before association rule analysis.

[0039] In the above implementation, to study the key risk factors affecting the degree of accident risk of autonomous vehicles and their coupling influence mechanisms, and to generate high-risk scenario profiles, an interpretable machine learning and association rule combination model is established, such as... Figure 1 As shown, the process begins with preprocessing accident data from autonomous vehicles. A predictive model for accident risk is then established using ROSE technology combined with the LightGBM machine learning algorithm. The SHAP interpretable module is extended to identify key risk factors. This model is then used to perform association rule analysis on high-risk accidents to observe multi-factor coupling patterns. Finally, a profile of high-risk accident scenarios is created.

[0040] In the above implementation, for future research on the accident risk levels of autonomous vehicles, the developed model should possess the ability to handle high-dimensional data input, training, and computation quickly and conveniently. For example, supervised machine learning models allow the model to quickly classify and predict the risk level of potential vehicle accidents based on a wealth of information such as the environment and vehicle motion. Compared to single machine learning models such as decision trees and linear models, ensemble machine learning algorithms such as LightGBM, by combining the prediction results of multiple base learners, can significantly improve prediction performance, robustness, and generalization ability, and are commonly used to solve various classification and regression problems.

[0041] This embodiment samples the LightGBM model, improving computational efficiency through histogram-based feature splitting and gradient-based one-sided sampling strategies. It also supports automatic feature selection and importance ranking, effectively handling high-dimensional, multi-class accident data while maintaining high accuracy. Compared to other ensemble models, LightGBM is parallelizable, converges quickly, and offers high prediction accuracy. The formula for LightGBM is as follows: MERGEFORMAT (1) In the formula: The labels indicate the risk level of accidents involving autonomous vehicles (divided into three categories: high risk, medium risk, and low risk). This indicates the model's prediction results. This is a loss function used to measure the true level of risk. and prediction results The differences between them; This is a regularization term used to constrain the base learner. The complexity is reduced to prevent overfitting.

[0042] In the above implementation, to address the issue of significant differences in the proportion of samples with different accident risk levels in the autonomous vehicle accident dataset, an objective balancing technique is employed to balance the dataset. Widely used methods include SMOTE sampling, ADASYN sampling, and ROSE sampling. However, SMOTE and ADASYN, based on the assumption of linear interpolation in the neighborhood, generate new sample points in the feature space, which is difficult to accurately reflect the true nonlinear structure of the feature distribution and is prone to introducing spurious samples when there are many noisy samples. In contrast, ROSE sampling randomly generates new samples in the feature space according to the probability distribution of minority class samples, improving sample balance while maintaining the data distribution characteristics. Furthermore, ROSE sampling can improve the model's discriminative ability and stability without significantly increasing noise risk, making it particularly suitable for accident datasets with limited sample size, high variable dimensionality, and complex class boundaries. Therefore, this embodiment selects ROSE sampling to address the data imbalance problem.

[0043] In summary, this embodiment constructs the ROSE-LightGBM model to explore the impact of scene factors on accident risk. Simultaneously, SHAP analysis is employed to improve model interpretability. SHAP analysis is an algorithm based on cooperative game theory used to explain the influence of factors on the dependent variable, approximating the output of the prediction model as the sum of the contributions of each factor. For autonomous vehicle accidents, scene factors... Shapley value The calculation formula is as follows: MERGEFORMAT (2) In the formula: F This is a set of factors related to accident scenarios involving autonomous vehicles. S A subset of scene factors; x Values ​​are assigned to the scene factors corresponding to a single accident record; For subset S The number of scene factors; For the prediction model in input x Only use subsets S The predicted value at that time; The weight term represents a subset among all possible permutations of features. S The probability of occurrence.

[0044] In the above implementation, this embodiment explores the potential coupling patterns among key risk factors in high-risk accidents involving autonomous vehicles based on association rules. Association rules calculate the interrelationships between influencing factors, determine their correlation, and then filter out associations with practical value, identifying their paths of action. This method can intuitively uncover the high-dimensional coupling effects of multiple factors without pre-set model assumptions, and identify combinations of key risk factors through quantitative indicators, providing strong support for the analysis of high-risk scenarios in autonomous vehicle accidents.

[0045] Let A and B be the item sets of scenario factors, representing different combinations of scenario factors for autonomous vehicle accidents. A complete association rule typically begins with " The rule is expressed in the form of "A = B = B", where A is the preceding term (triggering condition or premise) and B is the succeeding term (result or associated target). This rule does not involve causality, but rather reveals the phenomenon of "simultaneous existence" or "one object implying another object". Association rule "A = B ... "A pre-defined support, confidence, and lift requirement must be met. Support, confidence, and lift are three important parameters characterizing association rules. Support refers to the frequency of itemset A in all incident samples; confidence refers to the conditional probability that itemset B is also included in incident samples containing itemset A; lift is another indicator of the strength of an association rule, which is the ratio of the rule's confidence to the support of itemset B. Association rules that meet the pre-defined minimum support and minimum confidence are considered strong association rules. Furthermore, if Lift > 1, it indicates a positive correlation between itemset A and itemset B, and is called a valuable strong association rule. The parameter calculation formulas are as follows:" MERGEFORMAT (3) MERGEFORMAT (4) MERGEFORMAT (5) In the formula: Sup(A) The support of itemset A; N A for A The number of times an itemset appears in a transaction set; N The total number of transactions in the transaction set; For association rules Confidence level; for A and B The probability of both occurring; For itemsets A itemsets B The degree of improvement.

[0046] In this embodiment, the results are analyzed as follows, including: Model comparison and risk factor identification: In this embodiment, ROSE sampling technology is used in combination with LightGBM, Random Forest (RF), XGBoost and CatBoost to establish an autonomous vehicle accident risk factor identification model. Accuracy, precision, recall and F1 score are used as model evaluation indicators, and the model is evaluated based on 5-fold cross-validation.

[0047] As shown in Table 2, the ROSE-LightGBM model outperforms the ROSE-XGBoost, ROSE-RF, and ROSE-CatBoost models. Its accuracy is 0.74, a 7% improvement over the ROSE-XGBoost model; its precision is 0.65, approximately 8% improvement over the ROSE-XGBoost model; its recall is 0.69, approximately 28% improvement over the ROSE-CatBoost model; and its F1 score is 0.67, approximately 25% improvement over the ROSE-CatBoost model. Therefore, the ROSE-LightGBM model was selected for risk factor identification.

[0048] Table 2 Horizontal comparison Figure 2 The analysis results of accidents with different risk levels show that the same factors have significantly different effects in different accident categories. Specifically, (a) is the SHAP contribution map of high-risk factors, (b) is the SHAP contribution map of medium-risk factors, and (c) is the SHAP contribution map of low-risk factors. The number of collision locations is the most important feature, reflecting the direct impact of the physical complexity of the accident on the risk level. It is noteworthy that the vehicle model and year, along with the driving mode, show a significant differentiation in risk level. Early models (2020 and earlier) and autonomous driving modes have positive SHAP values, both indicating a significant increase in the probability of high-risk accidents. Existing research suggests that if autonomous driving trajectory planning algorithms do not fully consider human driving habits, the incidence of high-risk accidents may increase. In summary, the type of vehicle collision contributes significantly to high-risk accidents, season has a significant impact on medium-risk accidents, and the time of the accident has a significant effect on low-risk accidents. Specifically, rear-end collisions are more likely to lead to high-risk accidents; medium-risk accidents are more concentrated in winter; and the probability of low-risk accidents increases in the morning due to drivers being in better mental condition.

[0049] At the same time, by Figure 2 (a) It is evident that the top 10 features in high-risk accidents are significantly more important than the subsequent factors, while the importance of features after the 10th feature gradually decreases and tends to stabilize. Therefore, in this embodiment, the top 10 features are selected as key risk factors (number of collision locations, vehicle collision type, season, number of lanes, vehicle's pre-collision movement, collision object's pre-collision movement, driving mode, collision object type, accident time, and vehicle model year) for subsequent association rule mining analysis.

[0050] Furthermore, based on the key risk factors identified above, this embodiment uses association rules to explore the potential coupling patterns between key risk factors in high-risk accidents involving autonomous vehicles. The greater the lift of such association rules, the more significantly the leading term increases the probability of high-risk accidents occurring in the subsequent term, which has important significance for risk warning and accident interpretation.

[0051] To balance the number and strength of rules and identify strong association rules in high-risk incidents, based on existing research and through repeated experiments, this embodiment sets the minimum support threshold to 0.13 and the minimum confidence threshold to 0.7, and retains only valuable strong association rules with a lift greater than 1. Finally, based on the Apriori algorithm, a total of 557 valid rules were mined. Figure 3 The density distribution of strongly correlated rules for high-risk accidents involving autonomous vehicles is shown, making it easy to intuitively observe the number of rules and their attribute characteristics.

[0052] Depend on Figure 3 As can be seen, most rules are concentrated in the low support (0.15~0.20) and high confidence (0.70~0.80) range. Due to the diverse causes of traffic accidents, each factor appears relatively infrequently in the overall dataset, resulting in low support. Once specific antecedent conditions are met, the probability of the corresponding successor term appearing is high, thus leading to higher confidence, indicating that multiple factors interact during the accident process.

[0053] Furthermore, to visually demonstrate the potential coupling relationships among key risk factors in high-risk accidents involving autonomous vehicles, a network of key accident risk factors is constructed based on association rules to obtain nodes and their association relationships. This network is simplified to an undirected network containing 25 nodes and 99 edges. The node size reflects the frequency of occurrence of key risk factors, and the edge weights represent the lift of the association rules between key risk factors. The magnitude of the lift is reflected in the edge thickness within the network. Figure 4 As shown.

[0054] Depend on Figure 4It can be observed that factors such as autonomous driving mode, the collision object traveling straight, the vehicle reversing or parking, rear-end collision of the vehicle, and the collision object being a small vehicle are located at the core of the network, with high density and thick edges connecting them, indicating that they play a dominant role in high-risk accidents and form a close coupling relationship with various other factors, constituting the key feature combination of high-risk scenarios. Overall, the network exhibits characteristics of multi-factor interaction and a high concentration of core nodes, revealing the complexity and systemic nature of high-risk accidents involving autonomous vehicles.

[0055] To uncover more representative factor relationships, the top 5 strong association rules were selected from high to low lift for analysis, as detailed in Table 3.

[0056] Table 3 Table 3 can be used to summarize the "core risk characteristics combination". Coupling of Synergistic Factors The transmission path of "high-risk manifestation" is revealed. Among them, the association rule with the highest elevation shows that when the vehicle is reversing or stopped and the collision object is a small car, the straight-moving collision object is very likely to cause a rear-end collision with the vehicle, increasing the probability of a high-risk accident by 1.93 times. This indicates that the factor coupling mode in this leading term is the most dangerous accident precondition. Overall, there is a significant high-frequency co-occurrence among the factors of the vehicle reversing or stopping, the collision object being a small car, the rear-end collision of the vehicle, and the collision object moving straight, which is a core risk feature combination, consistent with the previous conclusions. Furthermore, when this combination works in conjunction with factors such as autonomous driving mode, nighttime or low-light periods, narrow road environments (less than 4 lanes), and early vehicle models, the corresponding elevation of the association rule indicates that the coupling of these factors will increase the probability of a high-risk accident by 1.87 to 1.92 times.

[0057] Based on the analysis results in Table 3, the following four typical high-risk scenarios for traffic accidents involving autonomous vehicles can be summarized: 1) Rear-end collision scenarios when the vehicle is reversing or stopped, such as Figure 5 As shown in (a), the vehicle's active collision avoidance capability decreases significantly when reversing or stationary. When the perception system misses or delays in recognizing small vehicles, especially when the other vehicle is traveling straight, a rear-end collision is highly likely to occur. In this scenario, the system response relies on passive braking, resulting in an extremely high risk of accident.

[0058] 2) Narrow road conflict scenarios in nighttime autonomous driving mode, such as Figure 5 As shown in (b), visual perception performance degrades under nighttime or low-light conditions, reducing the autonomous driving system's ability to detect and predict the trajectory of small vehicles. In narrow road environments, the system's collision avoidance decision space is limited, making it difficult to effectively respond to oncoming vehicles and increasing the risk of collision.

[0059] 3) Early models' self-reversing or stopping collision scenarios, such as Figure 5 As shown in (c), the perception algorithms and decision-making models of early models from 2020 and earlier are relatively simple, resulting in significant delays in response to sudden targets when reversing or stopped. The small number of collision locations indicates that accidents are mostly caused by a single key conflict, the system lacks redundant obstacle avoidance strategies, and its ability to identify and avoid small cars is insufficient, leading to an increased risk of rear-end collisions.

[0060] 4) Failure scenarios in multi-module collaboration for autonomous driving, such as Figure 5 As shown in (d), in autonomous driving mode, the system needs to coordinate the perception, decision-making, and control modules. When the vehicle is reversing or stopped, and the other vehicle is a small car traveling straight, there is a time synchronization problem and decision hesitation among the multiple modules, such as perception delay, trajectory prediction deviation, or braking command lag, which significantly increases the probability of a rear-end collision.

[0061] In conclusion: 1) Among the key risk factors, vehicle model and driving mode show a significant differentiation in risk level, with earlier models and autonomous driving modes both significantly increasing the probability of high-risk accidents. In addition, apart from the most significant factor of the number of collision locations, collision type, season, and accident time show strong predictive contributions in high-risk, medium-risk, and low-risk accident categories, respectively.

[0062] 2) In the high-risk accident scenario profile, features such as the vehicle reversing or stopping, the collision object traveling straight, and the collision object being a small car colliding with the rear of the vehicle exhibit a strong coupling pattern. When coupled with features such as autonomous driving mode, narrow road environment, and early vehicle models, the probability of high-risk accidents increases by 1.87 to 1.92 times.

[0063] 3) The following are four typical high-risk accident scenarios for autonomous vehicles: rear-end collisions when the vehicle is reversing or stopped, narrow road conflicts in nighttime autonomous driving mode, collisions when the vehicle is reversing or stopped due to system limitations of early models, and failures in the coordination of multiple autonomous driving modules.

[0064] 4) Currently, there is limited data available on domestic autonomous driving accidents. The next step should be to acquire more domestic autonomous vehicle accident data to conduct accident analysis in China. Simultaneously, by acquiring more detailed information such as the speed, acceleration (deceleration), and macroscopic traffic information of autonomous vehicles, we can explore the impact of more comprehensive single factors and their coupled effects on the degree of accident risk for autonomous vehicles.

[0065] Example 3 This embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method described in Embodiment 1.

[0066] Example 4 This embodiment also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0067] Example 5 This embodiment also discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in Embodiment 1.

[0068] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An automatic driving vehicle high-risk accident scene identification system, characterized in that, include: The data acquisition and preprocessing module is used to acquire accident data from autonomous vehicles and determine scenario factors; The risk factor identification module, connected to the data acquisition and preprocessing module, is used to receive the accident data and the scenario factors, establish a machine learning-based prediction model for the degree of accident risk of autonomous vehicles, and identify key risk factors. The association rule mining module is connected to the risk factor identification module and is used to receive the key risk factors, mine the coupling patterns between factors in high-risk accidents through association rule mining, and generate strong association rules. The scene profile generation module is connected to the association rule mining module. It is used to receive the strong association rules, construct a high-risk scene profile of autonomous vehicle accidents based on the coupling mode, and obtain the identification result of high-risk scene of autonomous vehicle accidents.

2. The system according to claim 1, characterized in that, The data acquisition and preprocessing module includes: The data cleaning unit is used to collect collision accident reports from autonomous vehicles and clean incomplete data. The feature expansion unit, connected to the data cleaning unit, is used to expand the technical feature variables of autonomous vehicles based on traditional scenario factors; the technical feature variables of autonomous vehicles include vehicle model year, sensor type and driving mode; The risk level classification unit, connected to the feature extension unit, is used to classify the degree of accident risk into three levels: high risk, medium risk, and low risk.

3. The system according to claim 1, characterized in that, The risk factor identification module includes: The data balancing unit is used to balance imbalanced accident data using a random oversampling instance method to generate balanced accident data. The model building unit, connected to the data balancing unit, is used to build an accident risk prediction model based on LightGBM based on the balanced accident data. The key factor identification unit, connected to the model building unit, is used to calculate the contribution value of each scenario factor to the output of the prediction model by combining the SHAP algorithm, and to select key risk factors according to the magnitude of the contribution value.

4. The system according to claim 1, characterized in that, The association rule mining module includes: The rule mining unit is used to mine association rules for key risk factors in high-risk accidents based on the Apriori algorithm. The rule filtering unit, connected to the rule mining unit, is used to filter strongly correlated rules that meet the preset minimum support, minimum confidence, and lift greater than 1. The network construction unit, connected to the rule filtering unit, is used to construct an accident key risk factor network with key risk factors as nodes and lift degree as edge weights based on the filtered strong correlation rules.

5. The system according to claim 1, characterized in that, The scene profile generation module includes: The feature combination extraction unit is used to extract the core risk feature combination and synergistic factors among the factors based on the rules with the highest lift in the strong association rules. The scenario summarization unit, connected to the feature combination extraction unit, is used to summarize and form a typical high-risk accident scenario profile based on the core risk feature combination and synergistic factors.

6. An automatic driving vehicle high-risk accident scene identification method, characterized in that, include: Acquire accident data from autonomous vehicles and determine scenario factors; Based on the accident data and scenario factors, a machine learning-based model for predicting the degree of accident risk of autonomous vehicles is established, and key risk factors are identified. Based on the key risk factors, strong association rules are generated by mining the coupling patterns between factors in high-risk accidents through association rule mining. Based on the strong association rules, a high-risk scenario profile for autonomous vehicle accidents is constructed, and the identification results of high-risk scenarios for autonomous vehicle accidents are obtained.

7. The method according to claim 6, characterized in that, The process of acquiring accident data from autonomous vehicles and identifying scenario factors includes: Collect collision accident reports from autonomous vehicles and clean incomplete data; In addition to traditional scenario factors, the technical characteristic variables of autonomous vehicles are expanded; among them, the technical characteristic variables of autonomous vehicles include vehicle model year, sensor type and driving mode; The degree of accident risk is divided into three levels: high risk, medium risk, and low risk.

8. The method according to claim 6, characterized in that, The process of establishing a machine learning-based model for predicting the level of accident risk for autonomous vehicles and identifying key risk factors includes: The imbalanced accident data is balanced using a random oversampling instance method to generate balanced accident data. Based on the balanced accident data, an accident risk prediction model based on LightGBM is constructed. The contribution value of each scenario factor to the prediction model output is calculated by combining the SHAP algorithm, and key risk factors are selected based on the magnitude of the contribution value.

9. The method according to claim 6, characterized in that, The process of generating strong association rules by mining coupling patterns among factors in high-risk incidents includes: Based on the Apriori algorithm, association rules are mined for key risk factors in high-risk accidents; Filter out strong association rules that meet the preset minimum support, minimum confidence, and lift greater than 1.

10. The method according to claim 6, characterized in that, After generating the strong association rule, the method further includes: Based on the selected strong correlation rules, an accident key risk factor network is constructed with the key risk factors as nodes and the lift degree as the edge weight.

Images