Intelligent driving risk scene classification and evaluation method and system based on traffic management accidents

Abstract

The application discloses an intelligent driving risk scene classification and evaluation method and system based on traffic management accidents, and belongs to the field of intelligent driving safety testing, and comprises the following steps: constructing an accident record set, guiding a large language model to extract structured parameter data of each accident record in the accident record set by using a prompt word; encoding participant types, conflict forms and road types based on the structured parameter data; constructing an accident causality vector based on the structured parameter data and the encoded data, classifying accident causes and effects according to the accident causality vector; analyzing the frequency, severity and predictability of each accident record in the accident record set according to the accident cause and effect classification result, and obtaining a comprehensive value index based on the frequency, severity and predictability; and sorting each accident record according to the comprehensive value index to obtain a value rating list. The application adopts the above method and system, and realizes text analysis, risk scene classification and three-dimensional comprehensive value evaluation.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving safety testing technology, and in particular to an intelligent driving risk scenario classification and assessment method and system based on traffic management accidents. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, autonomous driving has become an important strategic direction for the global automotive industry upgrade. However, intelligent driving systems still face severe safety risks and challenges in real, complex, and open road traffic environments. Ensuring the safe operation of intelligent driving vehicles in various traffic scenarios has become a core bottleneck restricting the large-scale commercialization of intelligent driving technology.

[0003] Scenario-based testing and verification methods have been widely recognized as an effective way to assess the safety of intelligent driving systems. This method verifies the safety and reliability of the system in complex environments by constructing test cases covering various driving scenarios. The International Organization for Standardization (ISO) has published a series of standards, including ISO 34501 "Road vehicles - Automated driving systems test scenarios - Vocabulary and terminology" and ISO 34502 "Road vehicles - Automated driving systems test scenarios - Scenario-based safety assessment framework." Domestically, corresponding standards are being developed and expanded to include commonly used terms in autonomous driving testing, such as "nominal scenario," "hazardous scenario," "edge scenario," "critical scenario," and "extreme scenario." Meanwhile, multiple institutions and research teams have conducted research and exploration on the classification and assessment of intelligent driving risk scenarios from different perspectives.

[0004] Existing methods for clustering accident data and classifying scenarios have the following problems: 1. Existing technologies primarily rely on predefined fields in structured incident databases such as CIDAS and NAIS for clustering, without involving unstructured text parsing. Current LLM-assisted incident text processing mainly transforms incident reports into logical scenarios through Intermediate Pattern Sequences (IPS), focusing on scenario generation rather than classification, sorting, and value assessment.

[0005] 2. There is a disconnect between scenario classification and risk assessment. The existing classification system classifies accidents by form (rear-end collision / side collision / intersection, etc.) without considering environmental factors. In the case of multiple factors coupled in "people-vehicle-road-environment", the classification results are difficult to support fine-grained risk assessment.

[0006] 3. Traditional scenario value assessment only considers frequency and severity, ignoring the unique dimension of intelligent driving: whether it can be predicted by the perception system. This makes it difficult for the scenario classification system to directly guide the generation of test cases and resource allocation. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for classifying and assessing intelligent driving risk scenarios based on traffic management accidents, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, this invention provides a method for classifying and assessing intelligent driving risk scenarios based on traffic management accidents, comprising the following steps: S1. Construct an accident record set, use prompt words to guide the large language model to extract the structured parameter data of each accident record in the accident record set, and verify the structured parameter data. The accident record includes accident identification text and structured accident record. S2. Based on structured parameter data, the PCR three-dimensional coding system is used to encode participant type, conflict pattern and road type; S3. Based on structured parameter data and coded data, construct accident causal vectors and classify accident causal events according to the accident causal vectors; S4. Based on the accident causal classification results, analyze the frequency, severity, and predictability of each accident record in the accident record set, and obtain a comprehensive value index based on the frequency, severity, and predictability. S5. Sort each accident record according to the comprehensive value index to obtain a value rating list.

[0009] Preferably, the accident determination text includes a description of the accident, a cause analysis, and a liability determination; the structured accident record includes the accident time and location, road type, lane / road geometry, participant type, weather / light / road surface condition, accident consequences, liability determination, and video evidence index; the structured parameters include the accident number and source type, traffic participant type, vehicle type, pre-accident behavior, illegal behavior, regulations or determination basis, liability allocation, conflict pattern, road type, environmental risk factors, vehicle technical failure, accident consequences, and sensor identifiable attributes.

[0010] Preferably, step S2 includes: extracting data related to traffic participant type, vehicle type, conflict pattern, and road type from the structured parameters, and encoding the participant type, conflict pattern, and road type using the PCR three-dimensional coding system.

[0011] Preferably, in the PCR three-dimensional coding system, the P code represents the participant type, including motor vehicle-motor vehicle, motor vehicle-non-motor vehicle, motor vehicle-pedestrian and single-vehicle accidents; the C code represents the conflict type, including road segment conflict-rear-end collision, road segment conflict-side collision, road segment conflict-frontal collision, intersection conflict-straight intersection, intersection conflict-left turn, intersection conflict-right turn and special area conflict; and the R code represents the road type, including highways, urban expressways, urban arterial roads, urban secondary arterial roads / branch roads, and national highways / provincial highways.

[0012] Preferably, step S2 further includes: mapping the PCR three-dimensional code to the ISO 34502 standard, where the P code corresponds to the participant type in ISO 34502, the C code corresponds to the functional scenario level, and the R code corresponds to the road attribute of the running design domain.

[0013] Preferably, step S3 specifically includes: Extract data related to illegal behavior, environmental risk factors, liability allocation, and vehicle technical malfunctions from structured parameters, and combine them with coded data to construct an accident causal vector. ,in, Indicates the severity of the participant's illegal behavior. Indicates environmental risk factors, Indicates a technical malfunction in the vehicle. Indicates the allocation of responsibilities; Cross-validation of the intensity of participants' illegal behavior, environmental risk factors, vehicle technical failures and liability allocation, and causal analysis statements in the accident report yields a classification of accident causality, including human-driven illegal behavior, environmental factors-driven, vehicle technical failure-driven, and mixed types. Among them, mixed types are marked with primary and secondary types or double-marked according to the liability ratio and causal analysis statements. When it is impossible to judge based on the liability ratio and causal analysis statements, manual review is output.

[0014] Preferably, step S4 specifically includes: For each PCR code k and its classification result, the corresponding accident subset is statistically analyzed; The formula for calculating the frequency of each incident record in the incident subset is: ; In the formula, For encoding The number of accidents, Total number of accidents; The severity of each incident record in the incident subset is calculated using the following formula: ; In the formula, The number of fatal accidents, The number of serious injury accidents, The number of minor injury accidents, , , All statistics are based on the consequences of the accidents; Analyze the predictability of each incident record in the incident subset; The comprehensive value index is derived based on frequency, severity, and predictability, using the following formula: ; In the formula, , , Represents the weight, satisfying , Representation encoding The predictability of the accident Represents the normalized frequency exponent. This represents the severity index after normalization.

[0015] Preferably, the predictability of each incident record in the incident subset is analyzed, including: Extract sensor-identifiable attributes, conflict patterns, environmental risk data, and vehicle technical fault-related data from structured parameters; Predictability is assessed based on the extracted data and the classification type: If detected by sensors ≥2 seconds before a collision, predictability This indicates that it is entirely predictable; If detected by sensors less than 2 seconds before a collision, predictability is high. This indicates that some aspects are foreseeable; If it is an environmental degradation scenario, predictability This indicates that some foreseeable scenarios of environmental degradation include rain, snow, fog, nighttime, and slippery conditions; If it cannot be detected by sensors at all, predictability This indicates that something is unpredictable.

[0016] Preferably, each incident record in the value rating list includes a code, scene name, frequency ranking, severity ranking, predictability, classification type label, comprehensive value index, value level, corresponding ISO 34502 number, and recommended test pass criteria.

[0017] This invention also provides an intelligent driving risk scenario classification and assessment system based on traffic management accidents, comprising: The accident record parsing module is used to construct an accident record set and use prompt words to guide the large language model to extract structured parameter data, and to verify the structured parameter data; The scene coding module is used to generate PCR 3D codes based on structured parameter data; The causal classification module is used to construct accident causal vectors based on structured parameter data and coded data, classify accident causal factors based on accident causal vectors, and output classification type labels. The value assessment module is used to calculate the frequency, severity, and predictability of each accident record in the accident record set based on the accident causal classification results, and to obtain a comprehensive value index. The list output module is used to sort each accident record according to the comprehensive value index and output a value rating list.

[0018] Therefore, the present invention employs the above-mentioned intelligent driving risk scenario classification and assessment method and system based on traffic management accidents, which has the following beneficial effects: (1) The input of this invention is accident identification text and structured accident record. It realizes the parsing of unstructured documents, and the accuracy of LLM structured parsing fields reaches 86.8%. It can realize the automatic conversion from millions of accident texts to a list with value rating. Compared with the existing manual accident scenario screening, it improves the processing efficiency.

[0019] (2) By combining environmental risk factors, this invention divides accidents into two categories: human-induced illegal activities and non-human-induced risk-induced scenarios. It provides differentiated pass criteria for regulatory compliance testing (zero collision) and risk scenario testing (collision mitigation), enabling the test plan to apply corresponding physical constraints to different types of scenarios.

[0020] (3) By introducing the predictability dimension, unpredictable extreme scenarios (such as road collapse) are not mistakenly given high testing priority, thus avoiding excessive allocation of testing resources to scenarios that the system cannot reasonably cope with. (4) The three-dimensional ranking based on frequency, severity and predictability provides greater objectivity.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a diagram illustrating the overall architecture of the intelligent driving risk scenario classification and assessment method based on traffic management accidents, as described in an embodiment of the present invention. Figure 2 This is a tree diagram of the three-dimensional coding system according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] Example like Figure 1As shown, this invention provides a method for classifying and assessing intelligent driving risk scenarios based on traffic accidents, including the following steps: S1. Construct an accident record set and input it into a large language model. Use prompt words to guide the large language model to extract structured parameter data for each accident record in the set, and validate the structured parameter data. The accident records include accident identification text and structured accident records. The large language model can be Tongyi Qianwen Qwen2.5-72B. Alternatively, a large language model with equivalent capabilities, such as GPT-4o, can be used; there are no specific restrictions on the large language model.

[0026] Specifically, the prompts use a three-part format: "role definition + output format constraints + few-shot example," as shown in the example below: Role Definition: You are a traffic accident structured analysis expert who needs to extract standardized structured information from accident reports.

[0027] Output format constraints: The output must strictly follow the JSON format, and the fields must include at least the accident number and source type, traffic participant type, vehicle type, pre-accident behavior, illegal behavior, legal or determination basis, liability allocation, conflict pattern, road type, environmental risk factors, vehicle technical failure, accident consequences, and sensor identifiable attributes.

[0028] Few-shot examples: Includes 5 annotation examples covering different accident types (rear-end collision, cutting-in collision, pedestrian crossing, intersection conflict, single-vehicle accident).

[0029] Accident determination text is typically 200-500 characters in natural Chinese, including a description of the accident, causal analysis, and liability determination. Structured accident records can be derived from CIDAS China Traffic Accident In-Depth Investigation Records, structured fields in the Public Security Traffic Management Accident Database, roadside equipment and test vehicle records from a regional intelligent connected vehicle test demonstration zone, and accident records with video IDs or sensor event IDs. Structured records typically include accident time and location, road type, lane / road geometry, participant type, weather / light / road surface condition, accident consequences, liability determination, and video evidence index. The output structured parameters include accident number and source type, traffic participant type, vehicle type, pre-accident behavior, illegal behavior, legal or determination basis, liability allocation, conflict pattern, road type, environmental risk factors, vehicle technical malfunction, accident consequences, and sensor identifiable attributes. Table 1 shows examples of structured output fields and values.

[0030] Table 1. Examples of Structured Output Fields and Values

[0031] Validating structured parameter data includes: JSON format verification: Check whether the output is valid JSON; Mandatory field check: At least one of the participant behaviors (road user type, vehicle type, pre-accident behavior, illegal act, regulation or identification basis) is present, and the conflict pattern and road type are not empty.

[0032] Logical consistency verification: For example, a "single-vehicle accident" shall not have more than 1 participant; the pre-accident behavior of the primarily responsible party in a "rear-end collision" shall include "going straight".

[0033] Uncertain field marking: Information not explicitly mentioned in the text (such as speed, weather) is marked as "not mentioned" instead of being guessed.

[0034] Cross-source consistency verification: When both text and structured fields exist for the same accident record, priority shall be given to the objective fields in the structured record (such as time, road type, weather, number of people affected by consequences), and the accident identification document text shall be used to supplement illegal acts, causal relationships and liability bases. If there is a conflict between the two sources, the corresponding field shall be marked as "conflict pending review".

[0035] The example of single text parsing is as follows: Input text: At about 2:30 on August 12, 2025, a person drove a small car with license plate number ChuanAXXXXX along the inner lane of the 3rd Ring Road in Chengdu from east to west. When driving to the downhill curve at Yangxi Flyover of the 3rd Ring Road, it happened to be raining and the road surface was slippery. The driver did not reduce the vehicle speed, causing the vehicle to lose control and hit the central isolation guardrail. This caused minor injuries to Li and damage to the vehicle. It is determined that Li shall bear full responsibility.

[0036] Example of synchronously input structured accident record: Road type = urban expressway; Road geometry = downhill curve; Weather = rainy; Road surface = slippery; Lighting = nighttime; Video record = available; Accident consequence = 1 person with minor injury; Liability determination = full liability of driver.

[0037] Extraction result: "Accident No.": "Example-001", "Input text": ["Accident identification document text", "Structured accident record of demonstration zone"], "Participant": ["Road user type": "motor vehicle", "vehicle type": "small car", "pre-accident behavior": "going straight - curve", "illegal act": "failing to reduce speed (failing to decelerate on slippery road)", "cause basis": "Cause analysis of accident identification document: failing to reduce speed on rainy and slippery road", "speed": "not mentioned"], "Conflict pattern": "single-vehicle accident - colliding with fixed object", "Road type": "urban expressway", "Road geometry": "Downhill curve" "Environmental Risk Factors": ["Weather": "Rainy", "Time": "Night", "Road Surface": "Slippery"], "Vehicle technical malfunction": "Not mentioned" "Sensor Recognizable Attributes": ["Obstruction": "No obvious obstruction", "Detectable Time Window": "Less than 2 seconds", "Time": "Nighttime"], "Accident Consequences": ["Severity": "Minor Injury", "Number of Deaths": 0, "Serious Injury Accident": 0, "Minor Injury Accident": 1], "Assignment of Responsibility":["Responsibility Determination":"Driver Li", "Label":"Full Responsibility", "Percentage":1.0"].

[0038] Performance metrics: Based on a validation set of 500 manually annotated accident reports, the classification accuracy of LLM (Qwen2.5-72B) is as follows: Conflict type: 87.4%; Road type: 91.2%; Illegal behavior identification: 85.6%; Overall field accuracy: 86.8%. This data is the validation result of one embodiment and can be adjusted according to the actual accident database and annotation standards.

[0039] S2. Based on structured parameter data, the PCR three-dimensional coding system is used to encode participant types, conflict patterns, and road types. Specifically, this includes: Data related to traffic participant type, vehicle type, conflict pattern, and road type were extracted from the structured parameters. The PCR three-dimensional coding system was used to encode participant type, conflict pattern, and road type. The P code represents participant type, and the relevant codes are shown in Table 2. The C code represents conflict pattern, and the relevant codes are shown in Table 3. The R code represents road pattern, and the relevant codes are shown in Table 4.

[0040] Table 2 Participant Type Coding

[0041] Table 3 Conflict Pattern Coding

[0042] Table 4 Road Type Codes

[0043] Combination coding rule: P code-C code-R code, for example, P2-C1a-R1 represents a rear-end collision between different types of motor vehicles in the same lane on a highway.

[0044] The PCR three-dimensional coding is mapped to the ISO 34502 standard. The P code corresponds to the participant type in ISO 34502, the C code corresponds to the functional scenario level, and the R code corresponds to the road attributes of the operational design domain. Following Annex D's verification methodology, the PCR coding is mapped to the ISO 34502 functional scenario framework, and the coverage rate is calculated. In one embodiment, the coverage target for Chinese traffic accident data can be set to no less than 95%. Uncovered accident types are categorized into three types according to JAMA: A, B, and C (Category A is covered, Category B requires expanded road features, and Category C is illegal / unforeseen events).

[0045] By being compatible with ISO 34502, the scenario list output by this invention can be directly used for: cross-OEM test scenario library interoperability, data support for regulatory test scenarios, and technical integration with international OEMs and Tier 1 suppliers. The three-dimensional coding system is as follows: Figure 2 As shown.

[0046] S3. Based on structured parameter data and coded data, construct accident causality vectors, and classify accident causality according to these vectors. Specifically, this includes: Extract data related to illegal behavior, environmental risk factors, liability allocation, and vehicle technical malfunctions from structured parameters, and combine them with coded data to construct an accident causal vector. ,in, Indicates the severity of the participant's illegal behavior. Indicates environmental risk factors, Indicates a technical malfunction in the vehicle. This indicates the allocation of responsibility. An exemplary quantitative rule for responsibility allocation is: full responsibility = 1.0, primary responsibility = 0.7, shared responsibility = 0.5, secondary responsibility = 0.3, no responsibility = 0. If the accident report provides a specific percentage, that percentage will be used; if only the wording of responsibility is given, the above mapping will be used.

[0047] By cross-validating the intensity of participants' illegal behavior, environmental risk factors, vehicle technical malfunctions and liability allocation, and the causal analysis statements in the accident report, a classification of accident causality was obtained, including human-induced illegal behavior as the primary cause, environmental factors as the primary cause, vehicle technical malfunctions as the primary cause, and a mixed cause. The classification rules are as follows: Human-caused illegality is the main factor: When the violation field contains high-confidence violations (such as speeding, running red lights, drunk driving, drug driving, driving against traffic, illegal lane changing, failure to yield, fatigued driving, and using a mobile phone), and the violation has a direct causal relationship with the accident, and is cross-validated by the liability allocation data as the violator being fully responsible, primarily responsible, or having a liability ratio ≥ 0.7, it is labeled as Type 1.

[0048] Environmental factors dominate: When environmental risk factors include rain, snow, fog, nighttime, slippery conditions, construction zones, low light intensity, and patchy fog, and the data on illegal acts is empty, not mentioned, or only shows minor lack of due diligence (e.g., failure to maintain a safe distance in severe weather or failure to reduce speed on slippery roads), and there is a direct causal relationship between the environmental risk and the accident, it is labeled as Type 2-E.

[0049] Vehicle technical malfunction is the primary cause: When the vehicle technical fault field includes vehicle technical faults such as brake failure, tire blowout, steering failure, power interruption, light failure, sensor obstruction, or vehicle structural failure, and the fault has a direct causal relationship with the occurrence of the accident, it is labeled as Type2-F. Type2-F and Type2-E are sub-labels of Type2 vehicle faults. Type1 indicates human-caused violations as the primary cause, and Type2 indicates non-human-caused risks as the primary cause.

[0050] Hybrid type: When at least two of the following three categories—human-caused violations, environmental risks, and vehicle technical malfunctions—are present simultaneously, the primary label is determined based on the liability ratio and the causal analysis statement: If the liability ratio corresponding to a high-confidence violation is ≥0.7, then Type1 is the primary label, and Type2-E or Type2-F is retained as a secondary label; if the liability ratio is within the equal liability range of 0.4-0.6, then both Type1 and Type2 labels are output; if the violation is merely a minor lack of care under adverse conditions, and the environment or vehicle malfunction is the primary triggering condition, then Type2 is the primary label. If a reliable judgment cannot be made from the text, "Manual review" is output.

[0051] The causal classification of this invention combines the subjective nature of the illegal act (intentional / negligent), the objective nature of the environmental risk, the suddenness of the vehicle's technical malfunction, and the causal strength of the accident result (primary responsibility / full responsibility / equal responsibility) to form a four-element determination model of "behavior-environment / malfunction-cause-responsibility". The physical meaning of this model is: The causative behaviors in Type 1 scenarios are behaviors that a qualified driver should not have committed. Therefore, intelligent driving systems should have zero-tolerance obstacle avoidance capabilities for such scenarios. In Type 2 scenarios, the causative conditions mainly stem from environmental risks or vehicle technical malfunctions, exceeding the fully controllable boundaries of a qualified driver or intelligent driving system under normal operating conditions. Therefore, the test guidelines allow collisions to occur but require mitigation of collision severity and require the system to provide degrade, warning, or safe stopping strategies.

[0052] The technical advantages of this classification method: Type 1 scenarios correspond to the mistakes that "qualified and attentive drivers" should not make in the ALKS regulations (UNECE R157)—mistakes that intelligent driving must be able to avoid.

[0053] Type 2 scenarios correspond to "reasonably foreseeable risks" in ISO 21448 SOTIF—intelligent driving should be able to avoid or mitigate collisions.

[0054] The Type2-F vehicle fault sub-label corresponds to the low-predictability risk of "external system or vehicle-specific technical faults". It should be primarily used for sensor-visible faults, vehicle instability detection, minimum risk strategies, and post-event data closed-loop analysis.

[0055] S4. Based on the accident causality classification results, analyze the frequency, severity, and predictability of each accident record in the accident record set, and derive a comprehensive value index based on frequency, severity, and predictability. Specifically, this includes: In practical implementation, it can access the public security traffic management accident database, the accident database of road traffic safety research institutions, or the accident event database of local demonstration zones; for example, approximately 2 million accident records from the past three years can be used as D. For each PCR code k and its Type sub-label, the corresponding accident subset is statistically analyzed.

[0056] The formula for calculating the frequency of each incident record in the incident subset is: ; In the formula, For encoding The number of accidents, This represents the total number of accidents. Therefore, It reflects the statistical exposure of a certain type of risk scenario in accident big data.

[0057] The severity of each incident record in the incident subset is calculated using the following formula: ; In the formula, The number of fatal accidents, The number of serious injury accidents, The number of minor injury accidents, , , All statistics are based on the aggregation of the accident consequence field within the accident subset.

[0058] Analyze the predictability of each incident record in the incident subset. This includes: Extract sensor-identifiable attributes, conflict patterns, environmental risk data, and vehicle technical fault-related data from structured parameters; Predictability is assessed based on the extracted data and the classification type: If the triggering conditions (such as the vehicle in front decelerating or a target cutting in from the side) can be detected by the sensors ≥2 seconds before the collision, the predictability is high. "Completely predictable" indicates that the scenario is entirely foreseeable. For example, mass-produced millimeter-wave radar + camera fusion solutions have a high detection rate for targets ahead under TTC conditions of more than 2 seconds, so such scenarios are "completely predictable" for intelligent driving systems.

[0059] If the triggering condition is detected by the sensor less than 2 seconds before the collision, predictability This indicates that the situation is partially foreseeable. For example, in a "ghost peek" scenario, a pedestrian or non-motorized vehicle suddenly appears from behind an obstacle, and the effective detection window of the sensor is typically less than 1.5 seconds. In this case, the system may detect the pedestrian, but the reaction time is insufficient to completely avoid a collision; this is considered "partially foreseeable."

[0060] If it is an environmental degradation scenario, predictability This indicates that some foreseeable environmental degradation scenarios include rain, snow, fog, nighttime, and slippery conditions.

[0061] If a mechanical failure is completely undetectable by sensors, such as road subsidence, tire blowout, or vehicle structural failure, predictability is low. This indicates that something is unpredictable.

[0062] The comprehensive value index is derived based on frequency, severity, and predictability, using the following formula: ; In the formula, , , The weights can be calibrated based on road type, speed range, test resource constraints, or expert-annotated samples to meet the following requirements. Preferred . Representation encoding The predictability of the accident Represents the normalized frequency exponent. This represents the severity index after normalization. , The results are obtained using maximum-minimum normalization, quantile normalization, or percentile normalization within the same accident record set, with values ​​ranging from [0,1]. When the sample size of an accident subset is lower than a preset threshold, smoothing correction is performed using adjacent PCR coding subsets or expert-annotated samples. Specifically, adjacent PCR coding subsets refer to accident subsets that differ from the target PCR coding in at most one dimension of the P-code, C-code, or R-code, and whose different dimensions belong to the same superordinate category or a preset adjacency relationship. The adjacency relationship is determined based on the similarity of participant types, the superordinate category of conflict patterns, and the continuity of road grade. Simply put, it refers to the "neighbor scenario" with the closest risk mechanism in the coding space, used for sample borrowing smoothing in small sample statistics.

[0063] S5. Sort each accident record in descending order according to the comprehensive value index, select the scenario codes of the top 15-20 in the comprehensive value index as "high-value test scenarios", and output them to obtain a value rating list. When the same PCR code exists in both Type1 and Type2 subclasses, retain the sorting results separately. The value rating list is shown in Table 5.

[0064] Table 5 Value Rating List

[0065] In one embodiment of the expert review sample, the method of the present invention has a consistency rate of ≥75% with the subjective ranking of the top 5 scenes in the comprehensive value index. At the same time, compared with the experts, it can identify high-frequency low-severity scenes that the experts did not notice (such as minor rear-end collisions related to AEB calibration).

[0066] This invention also provides an intelligent driving risk scenario classification and assessment system based on traffic management accidents, comprising: The accident record parsing module is used to construct an accident record set and use prompt words to guide the large language model to extract structured parameter data, and to verify the structured parameter data; The scene coding module is used to generate PCR 3D codes based on structured parameter data; The causal classification module is used to construct accident causal vectors based on structured parameter data and coded data, classify accident causal factors based on accident causal vectors, and output classification type labels. The value assessment module is used to calculate the frequency, severity, and predictability of each accident record in the accident record set based on the accident causal classification results, and to obtain a comprehensive value index. The list output module is used to sort each accident record according to the comprehensive value index and output a value rating list.

[0067] The present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor calls the computer program in the memory, it implements the content of the intelligent driving risk scenario classification and assessment method based on traffic management accidents.

[0068] The present invention also provides a storage medium storing computer-executable instructions, which, when loaded and executed by a processor, implement the content of an intelligent driving risk scenario classification and assessment method based on traffic management accidents.

[0069] Therefore, the present invention adopts the above-mentioned intelligent driving risk scenario classification and evaluation method and system based on traffic management accidents, realizing text parsing, risk scenario classification and three-dimensional comprehensive value evaluation, providing reliable scenario priority ranking and differentiated testing criteria for intelligent driving testing.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for classifying and assessing intelligent driving risk scenarios based on traffic management accidents, characterized in that, Includes the following steps: S1. Construct an accident record set, use prompt words to guide the large language model to extract the structured parameter data of each accident record in the accident record set, and verify the structured parameter data. The accident record includes accident identification text and structured accident record. S2. Based on structured parameter data, the PCR three-dimensional coding system is used to encode participant type, conflict pattern and road type; S3. Based on structured parameter data and coded data, construct accident causal vectors and classify accident causal events according to the accident causal vectors; S4. Based on the accident causal classification results, analyze the frequency, severity, and predictability of each accident record in the accident record set, and obtain a comprehensive value index based on the frequency, severity, and predictability. S5. Sort each accident record according to the comprehensive value index to obtain a value rating list.

2. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 1, characterized in that: The accident determination text includes a description of the accident, a cause analysis, and a liability determination; the structured accident record includes the accident time and location, road type, lane / road geometry, participant type, weather / light / road surface condition, accident consequences, liability determination, and video evidence index; the structured parameters include the accident number and source type, traffic participant type, vehicle type, pre-accident behavior, illegal behavior, legal or determination basis, liability allocation, conflict pattern, road type, environmental risk factors, vehicle technical failure, accident consequences, and sensor identifiable attributes.

3. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 2, characterized in that, Step S2 includes: extracting data related to traffic participant type, vehicle type, conflict pattern, and road type from the structured parameters, and encoding the participant type, conflict pattern, and road type using the PCR three-dimensional coding system.

4. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 3, characterized in that: In the PCR three-dimensional coding system, P code represents the participant type, including motor vehicle-motor vehicle, motor vehicle-non-motor vehicle, motor vehicle-pedestrian and single-vehicle accidents; C code represents the conflict type, including road segment conflict-rear-end collision, road segment conflict-side collision, road segment conflict-frontal collision, intersection conflict-straight intersection, intersection conflict-left turn, intersection conflict-right turn and special area conflict; R code represents the road type, including highway, urban expressway, urban arterial road, urban secondary arterial road / branch road, national highway / provincial highway.

5. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 4, characterized in that, Step S2 also includes mapping the PCR three-dimensional code to the ISO 34502 standard, where the P code corresponds to the participant type in ISO 34502, the C code corresponds to the functional scenario level, and the R code corresponds to the road attribute of the running design domain.

6. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 5, characterized in that, Step S3 specifically includes: Extract data related to illegal behavior, environmental risk factors, liability allocation, and vehicle technical malfunctions from structured parameters, and combine them with coded data to construct an accident causal vector. ,in, Indicates the severity of the participant's illegal behavior. Indicates environmental risk factors, Indicates a technical malfunction in the vehicle. Indicates the allocation of responsibilities; Cross-validation of the intensity of participants' illegal behavior, environmental risk factors, vehicle technical failures and liability allocation, and causal analysis statements in the accident report yields a classification of accident causality, including human-driven illegal behavior, environmental factors-driven, vehicle technical failure-driven, and mixed types. Among them, mixed types are marked with primary and secondary types or double-marked according to the liability ratio and causal analysis statements. When it is impossible to judge based on the liability ratio and causal analysis statements, manual review is output.

7. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 6, characterized in that, Step S4 specifically includes: For each PCR code k and its classification result, the corresponding accident subset is statistically analyzed; The formula for calculating the frequency of each incident record in the incident subset is: ； In the formula, For encoding The number of accidents, Total number of accidents; The severity of each incident record in the incident subset is calculated using the following formula: ； In the formula, The number of fatal accidents, The number of serious injury accidents, The number of minor injury accidents, , , All statistics are based on the consequences of the accidents; Analyze the predictability of each incident record in the incident subset; The comprehensive value index is derived based on frequency, severity, and predictability, using the following formula: ； In the formula, , , Represents the weight, satisfying , Representation encoding The predictability of the accident Represents the normalized frequency exponent. This represents the severity index after normalization.

8. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 7, characterized in that, The predictability of each incident record in the incident subset is analyzed, including: Extract sensor-identifiable attributes, conflict patterns, environmental risk data, and vehicle technical fault-related data from structured parameters; Predictability is assessed based on the extracted data and the classification type: If detected by sensors ≥2 seconds before a collision, predictability This indicates that it is entirely predictable; If detected by sensors less than 2 seconds before a collision, predictability is high. This indicates that some aspects are foreseeable. If it is an environmental degradation scenario, predictability This indicates that some foreseeable scenarios of environmental degradation include rain, snow, fog, nighttime, and slippery conditions. If it cannot be detected by sensors at all, predictability This indicates that something is unpredictable.

9. The intelligent driving risk scenario classification and assessment method based on traffic management accidents according to claim 8, characterized in that: Each incident record in the value rating list includes a code, scene name, frequency ranking, severity ranking, predictability, classification type label, comprehensive value index, value level, corresponding ISO 34502 number, and recommended test pass criteria.

10. A system for classifying and assessing intelligent driving risk scenarios based on traffic management accidents, used to execute the method for classifying and assessing intelligent driving risk scenarios based on traffic management accidents as described in any one of claims 1-9, characterized in that, include: The accident record parsing module is used to construct an accident record set and use prompt words to guide the large language model to extract structured parameter data, and to verify the structured parameter data; The scene coding module is used to generate PCR 3D codes based on structured parameter data; The causal classification module is used to construct accident causal vectors based on structured parameter data and coded data, classify accident causal factors based on accident causal vectors, and output classification type labels. The value assessment module is used to calculate the frequency, severity, and predictability of each accident record in the accident record set based on the accident causal classification results, and to obtain a comprehensive value index. The list output module is used to sort each accident record according to the comprehensive value index and output a value rating list.

Images