A method for assisting in determining liability in traffic accidents based on a large language model

Abstract

This invention discloses a traffic accident liability determination method based on a large language model, comprising: collecting data information on the vehicle's driving process before the accident; integrating and transmitting data information from various data sources, preprocessing the data information to obtain structured data, and generating a driving record description; calling a large language model to analyze the accident situation based on the structured data and the driving record description; generating a structured liability determination report based on the accident situation analysis of the vehicle, and transmitting the structured liability determination report to the traffic management department to assist in traffic accident liability determination. This invention utilizes the reasoning ability of a large language model to improve the efficiency of accident liability determination, reduce the time loss of multiple parties involved in the accident, reduce the workload of traffic management departments, speed up the insurance process, and improve the accuracy of accident liability determination.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent determination of vehicle traffic accidents, specifically relating to an auxiliary method for determining liability in traffic accidents based on a large language model. Background Technology

[0002] With the rapid development of cities and the advancement of autonomous driving technology, in the future information environment, vehicle-to-everything (V2X) technology and autonomous driving technology will mature, and V2X technology will be basically realized, enabling vehicles to communicate with each other quickly and obtain each other's information (relative speed, acceleration, location information, etc.).

[0003] Against the backdrop of my country's rapid economic development, the number of motor vehicles has continued to increase. According to statistics from the Ministry of Public Security, the number of motor vehicles in China reached 435 million in 2023, with 24.8 million new vehicles registered, an increase of 16,000 compared to 2022. With the increase in the number of vehicles, traffic accidents have also occurred frequently.

[0004] The traffic accident liability determination process includes multiple steps such as setting up safety warnings, reporting to the police, sending the injured to the hospital, taking photos for evidence, clearing the scene, and the final determination of liability. However, traditional traffic accident liability determination methods suffer from several bottlenecks. First, the on-site handling time for minor accidents is too long. For minor accidents, the processing time is usually at least 20 minutes. These minor accidents are one of the main causes of urban traffic congestion and may even trigger secondary accidents. Second, the liability determination cycle for non-minor accidents (including general accidents, major accidents, and particularly serious accidents) is long, and the damage assessment process is cumbersome. Traditional manual liability determination is slow and the damage assessment process is cumbersome, usually requiring several working days. Finally, accident liability determination consumes a lot of police resources. A large number of minor accidents may lead to a shortage of police resources, thereby affecting the allocation of police resources and reducing the efficiency of handling the situation.

[0005] The prior art relates to a method, apparatus, device, and readable storage medium for determining liability in autonomous driving accidents involving intelligent connected vehicles (CN116244664B). The liability determination method includes the following steps: S1, acquiring accident data of the autonomous driving accident; S2, determining whether the autonomous driving setting conditions are met; if yes, proceed to step S3; otherwise, proceed to step S6; S3, determining whether there was a malfunction in the ADS (Autonomous Driving System) before the autonomous driving accident occurred; if yes, proceed to step S5; otherwise, proceed to step S4; S4, determining whether the human intervention vehicle operation setting conditions are met; if yes, proceed to step S6; otherwise, proceed to step S5; S5, determining the vehicle manufacturer's liability and proceeding to step S7; S6, determining the driver's liability; S7, ending.

[0006] Its drawback lies in the fact that the patent's implementation method addresses the issue of liability after an autonomous driving accident. It determines whether the accident was caused by the driver or the autonomous driving of the connected vehicle, without assigning responsibility to all parties involved in the traffic accident. Furthermore, the patent's implementation method does not provide detailed liability assessment for traffic accidents and does not automatically generate an auxiliary liability assessment report. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned shortcomings in the prior art by providing a traffic accident liability assistance method based on a large language model, in order to solve the problem that existing traffic accident liability determination methods do not identify the responsibilities of multiple parties involved in the traffic accident.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A traffic accident liability determination method based on a large language model includes the following steps:

[0010] S1. Collect data information about the vehicle's driving process before the accident occurs;

[0011] S2. Integrate and transmit data information from various data sources, preprocess the data information to obtain structured data, and generate driving record descriptions;

[0012] S3. Based on the structured data and driving record descriptions, call the large language model to analyze the accident situation;

[0013] S4. The large language model generates a structured liability assessment report based on the analysis of the vehicle's accident situation, and transmits the structured liability assessment report to the traffic management department to assist in the determination of liability for traffic accidents.

[0014] Furthermore, the data information in S1 includes:

[0015] Vehicle information includes basic vehicle operation information, vehicle operating status information, and safety device status information;

[0016] Road and environmental data, including road conditions, traffic signals, and weather data;

[0017] External vehicle and environmental information, including the driving behavior of other vehicles and surrounding obstacles;

[0018] Furthermore, the preprocessing of data information in S2 includes:

[0019] Remove noise from sensor data, fill in missing data, and correct abnormal data;

[0020] Convert all data into a unified coordinate system and timestamp;

[0021] External data such as weather and road conditions are transformed into descriptions with contextual semantics through feature extraction;

[0022] The preprocessed structured data is transmitted to the large language model in real time and a driving record description is generated.

[0023] Furthermore, step S3 includes the following sub-steps:

[0024] S31. Before the accident, the large language model is called through the API interface, and the structured data is input into the large language model to generate a driving record description;

[0025] S32. During the accident occurrence phase, the large language model determines whether the current vehicle has experienced a sudden event based on the driving record description, and then determines whether a traffic accident has occurred.

[0026] S33. In the post-accident phase, the large language model generates descriptions and key memories with new driving record description frequencies.

[0027] S34. The large language model generates an accident snapshot based on key memories described in different driving records at three time stages: before, during, and after the accident.

[0028] Furthermore, S31 specifically includes:

[0029] The vehicle behavior description function is activated, and the structured data is input into the large language model to obtain the model's response content using the prompt word template of the large language model;

[0030] The model's response content is converted into a time-series format description of driving records;

[0031] Based on the current time series format of the driving record description and real-time data, generate a phase summary for this time period and store it as "Key Memory - Phase Driving Record Data Summary" as the new context input for the next time series.

[0032] Furthermore, step S32 includes the following sub-steps:

[0033] S321. Based on the "Key Memory - Summary of Stage Driving Data Records" in S31 as the new context input of the large language model, the large language model identifies whether there is a sudden event in the structured description; if no sudden event occurs, the stage description of the time series is stored as "Key Memory - Driving Record Description", and the vehicle speed, acceleration, and vehicle start-up status parameters are used to determine whether the vehicle has stopped.

[0034] If the vehicle stops, the large language model is invoked to summarize the trip information, generate a "key memory - summary of stage driving data records" and end the traffic accident judgment process;

[0035] If the vehicle does not stop, delete the data related to this time series from the large model context input, and use the staged descriptive memory as the context input for the next time series;

[0036] In the event of an emergency, proceed to S322;

[0037] S322. If a vehicle experiences a sudden event, the large language model determines whether a traffic accident has occurred and generates an analysis record.

[0038] If no traffic accident occurs, the characteristics of the emergency are described and analyzed and stored as key memories, and it is determined whether the vehicle has stopped. If the vehicle has stopped, the large language model summarizes the trip information based on all the stored key memories and ends the traffic accident judgment process.

[0039] If the vehicle does not stop, during the generation of the driving record description for the next time series, the data input from the previous time series will be deleted, and the key memory—the description and analysis record of the sudden event and the key memory—the summary of the driving record data in stages will be used as the new context input for the next time series.

[0040] If a traffic accident occurs, proceed to S333;

[0041] S323. The large language model increases the frequency of driving record descriptions and generates driving record descriptions as new contextual inputs for the large language model.

[0042] S324. Large language model determines whether a traffic accident has ended;

[0043] If the traffic accident is not yet over, a driving record description and a key memory—a summary of the driving record data for each stage are generated and stored.

[0044] If the traffic accident is resolved, proceed to S33.

[0045] Furthermore, S34 specifically includes:

[0046] The stored key memories, including the feature descriptions and analysis records of the sudden event and the summary of the driving record data of the previous stage before the accident, are used as new contextual inputs to the large language model to generate an accident snapshot, so as to provide an overview of the entire accident process.

[0047] Furthermore, S4 specifically includes:

[0048] Using the driving record data logs generated during the accident as contextual input, a large language model is used to visualize the data, generating curves and charts that change over time. The analysis of accident liability is performed by combining key memories from before, during, and after the accident, driving record descriptions, and accident snapshots, generating a structured liability report.

[0049] Furthermore, the time-varying curves include: a curve showing the change in speed over time, a curve showing the change in acceleration over time, and a curve showing the change in the relative distance between the vehicle in front and the vehicle behind.

[0050] The traffic accident liability determination method based on a large language model provided by this invention has the following beneficial effects:

[0051] 1. This invention improves the efficiency of accident liability determination, reduces the time loss of multiple parties involved in an accident (participants, insurance companies, traffic police), reduces the workload of traffic management departments, speeds up the insurance process, and improves the accuracy of accident liability determination.

[0052] 2. This invention utilizes the reasoning capabilities of a large language model to generate driving record descriptions. After a sudden event or traffic accident, it instantly generates driving record descriptions and structured examples of auxiliary accident liability determination. Based on collected real-time data and continuous driving record descriptions, it generates time-series-based visual charts displaying traffic accident liability determination methods, including speed, acceleration, and distance. The generated detailed and accurate driving record (accident) descriptions significantly reduce the time spent by traffic management departments and insurance companies in liability determination, providing accurate and detailed assistance to the accident determination parties, thereby reducing time losses for the accident parties, traffic management departments, and insurance companies. Real-time handling of traffic accidents can significantly reduce road congestion caused by accident handling and the potential risk of accidents due to road congestion.

[0053] 3. This invention utilizes a large language model to process and determine whether a sudden event or traffic accident has occurred, whether the accident has ended, and whether the vehicle has stopped, based on the collected information. It also generates time-series-based accident snapshots, structured liability assessment reports, and curves showing speed, acceleration, and relative distance to other vehicles over time to assist vehicle drivers, insurance companies, and traffic management departments in determining and assessing liability. Attached Figure Description

[0054] Figure 1 This is a flowchart of the traffic accident liability determination method based on a large language model according to the present invention.

[0055] Figure 2 This is an example of a prompt word template for generating driving records according to the present invention.

[0056] Figure 3This invention describes the time-based vehicle record description generation process.

[0057] Figure 4 This is a flowchart of the traffic accident judgment process of the present invention.

[0058] Figure 5 This is a snapshot case of an accident in this invention.

[0059] Figure 6 This is an example of a graph showing how the present invention changes over time.

[0060] Figure 7 This is an example of the accident liability determination result of the present invention. Detailed Implementation

[0061] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0062] Definitions of abbreviations and key terms in this invention

[0063] LLMs (Large Language Models) are deep learning-based Natural Language Processing (NLP) models that understand, generate, and predict language by training on large amounts of text data. LLMs can handle and generate complex language tasks such as text generation, translation, summarization, question answering, and sentiment analysis. Through large-scale parameters and training data, they capture deep-seated patterns and contextual relationships in language.

[0064] Time series: A time series is a set of data points arranged in chronological order. These data are typically collected over specific time intervals and possess the attribute of temporal sequence. In a time series, each data point represents a value observed at a specific moment or within a certain time period, and is usually closely related to a time variable.

[0065] API: Application Programming Interface. It is a set of definitions, protocols, and tools used to build software and applications, specifying how software components interact, including communication between libraries, operating systems, and microservices.

[0066] In the field of large language models, a "prompt" refers to the input text or instruction given by the user, which guides the model to generate the corresponding output. When users interact with a large language model, they may ask a question, a sentence, a paragraph, or any type of text content, all of which are called prompts or prompt words.

[0067] Key memory refers to the information that a language model retains during the current dialogue in a text-generating type of language model. This information helps in understanding and generating user interactions, enabling it to provide more context-aware, continuous, and relevant responses.

[0068] ABS: Anti-lock Braking System is a car safety system used to prevent wheel lock-up during emergency braking, thereby maintaining vehicle handling and stability.

[0069] Example 1

[0070] This embodiment presents a traffic accident liability assistance method based on a large language model, which has four processing stages and three time stages.

[0071] The three time phases include: the pre-accident phase, the accident phase, and the post-accident phase;

[0072] The four processing stages include:

[0073] The first stage is the information gathering stage; this stage corresponds to the pre-accident stage.

[0074] The second stage is the data transmission and preprocessing stage; this stage corresponds to the stage before the accident occurs.

[0075] The third stage is the accident situation analysis stage, which corresponds to the accident occurrence stage.

[0076] The fourth stage is the determination of responsibility for the accident and the output of results. This stage corresponds to the post-accident stage.

[0077] Based on this, refer to Figure 1 and Figure 4 The present invention specifically includes the following:

[0078] Step S1, Information Collection Phase; Collecting data on the vehicle's movement before the accident, specifically including the following:

[0079] Vehicle information refers to various information about the vehicle currently being driven by the driver, including basic vehicle operating information, vehicle operating status information, and safety device status information.

[0080] The basic information about the vehicle's current operating status includes time, location, speed, acceleration, vehicle steering angle, vehicle ID, and driver information.

[0081] Vehicle safety status information refers to the activation status of the vehicle's passive safety devices, such as airbags and ABS (anti-lock braking system). This information can be obtained through the vehicle's own acceleration sensors, speed sensors, and other sensors to determine the vehicle's current operating and safety status, helping to assess the severity of any potential accident.

[0082] Road and environmental data refers to road conditions, traffic signals, and weather data, including current weather conditions such as rain, snow, slippery roads, temperature, and wind speed. This data can be acquired through video cameras, radar detectors, and infrared sensors around the vehicle to monitor the vehicle's surrounding environment and detect vehicle safety in real time.

[0083] External vehicle and environmental information refers to the driving behavior of other vehicles besides the current vehicle. This is obtained through sensors such as radar and cameras, which acquire data such as the speed, acceleration, distance, relative position, and video data of the front, rear, left, and right sides of the vehicle. It also includes information about surrounding obstacles, such as whether there are pedestrians or obstacles around the vehicle, which is acquired using devices such as ultrasonic sensors and radar.

[0084] The specific information collected is shown in Table 1;

[0085] Table 1 Information Collection Form

[0086]

[0087]

[0088] Step S2, Data Transmission and Preprocessing Stage: This stage integrates and transmits data from various data sources, preprocesses the data to obtain structured data, and generates a vehicle record description. Specifically, it includes the following:

[0089] The main purpose of this stage is to integrate and transmit data from various data sources, including real-time information from other vehicles, environmental data obtained from API interfaces, and data from the current vehicle's sensors, and to clean, standardize, and preprocess this raw data. The preprocessed data will be transformed into structured text data that can be used by a large language model to generate accurate and real-time driving record descriptions.

[0090] During the data integration and transmission phase, information from various data sources is collected in real time and transmitted uniformly to the processing platform. This data includes information from vehicle sensors (such as GPS, IMU, and vehicle speed sensors), the surrounding traffic environment (such as other vehicles, traffic signals, and road conditions), and the external environment (such as weather information and traffic flow). Vehicles exchange information in real time with other vehicles and road infrastructure through onboard communication systems (such as CAN bus and V2X communication), ensuring that all relevant data can be transmitted to the system in a timely and accurate manner. Efficient transmission protocols and low-latency communication technologies ensure that the large language model can acquire complete driving data in real time.

[0091] In the data preprocessing stage, the raw data undergoes cleaning, standardization, and formatting to transform it into structured text data usable by a large language model. First, noise is removed from the sensor data, missing data is filled in, and outlier corrections are performed. Then, the data is standardized to a standard format; for example, vehicle speed, acceleration, and position data are converted into a unified coordinate system and timestamps. Next, feature extraction transforms external data such as weather and road conditions into contextualized descriptions. Finally, the processed structured data is transmitted to the language model in real time, enabling it to generate accurate driving record descriptions before, during, and after an accident.

[0092] Step S3, the accident situation analysis stage, is used to analyze the accident situation by calling a large language model based on structured data and driving record descriptions. The language model called via API here is a commonly used domestic language model (such as Tongyi Qianwen, Wenxin Yiyan, DeepSeekV3, etc.). After data transmission and preprocessing, processed structured data is obtained. To ensure data privacy and security and prevent the risk of data leakage, the proposed auxiliary liability determination method calls the large language model via API. No data is transmitted to the cloud during the entire auxiliary liability determination process; everything is integrated within the driver's current vehicle. Specifically, this includes three time periods: before the accident, during the accident, and after the accident. Based on the various information collected and processed in the previous stage, the large language model generates different reflections and response results, which specifically include the following sub-steps:

[0093] Step S31: Before the incident occurs, inference is performed by calling the large language model service via the API interface. Specifically, a pre-defined Prompt template is used, combined with a customized program, to input pre-processed structured information into the language model. (Reference) Figure 2 The language model prompt template used covers the task requirements, data structure, and data items represented by multiple structured data variables. The implementation steps are as follows: First, start the vehicle behavior description function; second, input the preprocessed structured data; then, use the prompt word template of the large language model to obtain the model's response content.

[0094] refer to Figure 3 The system transforms the response results of a large language model into time-series format dash dash descriptions. The dash dash description text is generated based on time-series data, continuously updating the text description according to changes in the collected real-time data. Before the accident, each time series lasts for T minutes, and the dash dash description content is updated every a1 seconds.

[0095] Time-series-based vehicle record description is an ongoing process, covering all time periods before, during, and after an accident. Next, the system proceeds to the next step: based on the real-time data and vehicle record description of the current time series (within T minutes), it generates a phased summary for that time period. This avoids excessive contextual input to the language model, preventing information overload, and ensures the description is structured and concise. The phased summary outlines the changes in the vehicle's operating status within that time period. This summary serves as an overview of the time period, helping the large language model retain key information in subsequent analyses and is stored as "Key Memory – Phased Vehicle Record Data Summary," serving as new contextual input for the next time series.

[0096] Step S32: During the accident occurrence phase, the large language model determines whether a sudden event has occurred in the current vehicle based on the driving record description, and then determines whether a traffic accident has occurred. This specifically includes the following sub-steps:

[0097] Step S321: Using the "Key Memory - Summary of Stage Driving Data Records" from step S31 as the new context input for the large language model, the large language model is required to identify whether there are any sudden events in the structured description, such as sudden speed changes, sudden acceleration drops, airbag deployment, anti-lock braking system (ABS) locking, or the onset of rain. If no sudden event occurs, the stage description of the time series is stored as "Key Memory - Driving Record Description," and the vehicle's stopping status is determined using parameters such as vehicle speed, acceleration, and vehicle activation status. If the vehicle stops, the language model is invoked and a new task is assigned—based on all stored key memories, the large model is required to summarize the trip information, generate "Key Memory - Summary of Stage Driving Data Records," and the traffic accident judgment process ends. If the vehicle does not stop, the data related to this time series in the large model's context input is deleted, and the stage description memory is used as the context input for the next time series, and the system continues to loop.

[0098] In the event of an emergency, proceed to S322;

[0099] Step S322: If a sudden event occurs to the vehicle, the large language model determines whether a traffic accident has occurred and generates an analysis record;

[0100] Specifically, if a vehicle encounters a sudden event, a new task is assigned to the large language model, requiring it to perform a sudden event feature analysis based on the key memories (summary of phased driving data records) from the previous step and the vehicle's structured data. The analysis uses acceleration thresholds, collision force magnitude, installation device status, and speed changes as accident judgment indicators to determine whether a traffic accident has occurred and generates an analysis record. If the analysis results indicate that no accident has occurred, the sudden event feature description and analysis record are stored as key memories (sudden event description and analysis), and parameters such as vehicle speed, acceleration, and activation status are used to determine whether the vehicle has stopped. If the vehicle has stopped, the large language model is invoked and a new task is assigned—requiring the large language model to summarize the trip information based on all stored key memories—and the traffic accident judgment process ends. If the vehicle has not stopped, during the generation of the driving record description for the next time series, the data input from the previous time series is deleted, and the key memories—sudden event description and analysis record and key memories—summary of phased driving record data—are used as new context inputs for the next time series, and the system continues the loop.

[0101] If a traffic accident occurs, proceed to step S333;

[0102] Step S333: The large language model increases the frequency of driving record descriptions and generates driving record descriptions as new context inputs for the large language model;

[0103] When a traffic accident occurs, the system assigns a new task—requiring the large language model to increase the frequency of driving record descriptions to capture more detailed accident information and generate richer driving record descriptions as new contextual input for the large language model. Specifically, the system generates driving record descriptions at a frequency that changes from a1 to a2, where a2 > a1. After increasing the driving record description frequency, the system proceeds to the next step.

[0104] Step S334: The large language model determines whether the traffic accident has ended;

[0105] If the traffic accident is not yet over, a driving record description and a key memory—a summary of the driving record data for each stage are generated and stored.

[0106] After adding the driving record description to the system, a new task is updated in the large language model—determining whether the accident has ended, using indicators such as whether the vehicle has stopped, whether the speed and acceleration have changed, and the vehicle's on / off status. If the accident is not yet over, the loop continues, generating the driving record description and key memory—a summary of the stage driving record data, storing only the key memory—the summary of the stage driving record data; if the accident is over, the system enters the post-accident stage and proceeds to step S33.

[0107] Step S33: In the post-accident phase, the large language model generates descriptions and key memories with new driving record description frequencies;

[0108] Specifically, after an accident is detected, the large language model is updated with a new task—requiring it to continue monitoring vehicle status (e.g., whether hazard lights are activated after parking, whether traffic has resumed) and generating a new description according to the new driving record description frequency a3. Since vehicles are often stationary after an accident, and vehicle data characteristics do not change significantly, to avoid information overload caused by a large driving record description frequency input to the language model, the new driving record description frequency a3 is lower than the initial driving record description frequency a1. After each time series, the system generates and stores a summary of the stage driving record data for subsequent generation and analysis. After generating descriptions and key memories with the new driving record description frequency, the process proceeds to step S34.

[0109] Step S34: The large language model generates an accident snapshot based on the key memories described in the different driving records at three time stages: before the accident, during the accident, and after the accident.

[0110] Specifically, based on key memories from different driving record descriptions at three time stages—before, during, and after the accident—an accident snapshot is generated. This snapshot provides an overview of key times and vehicle behavior before, during, and after the accident, aiding the large language model in subsequent accident liability analysis and responsibility determination. In other words, previously stored key memories, including feature descriptions and analysis records of the sudden event and summaries of driving record data from the previous stage before the accident, are used as new contextual input to the large language model to generate the accident snapshot. This snapshot provides an overview of the entire accident process, helping the language model understand the scenario and determine responsibility in subsequent scenarios.

[0111] refer to Figure 5 The accident snapshot includes a timestamp, key memories generated by the large language model in the previous step—a phased summary of driving description, vehicle operation information, environmental information, safety device activation status, event type judgment results generated by the large language model based on data input, accident status judgment results, and multiple information descriptions based on time changes.

[0112] Step S4: Based on the analysis of the vehicle's accident situation, the large language model generates a structured liability assessment report and transmits the structured liability assessment report to the traffic management department to assist in determining liability for traffic accidents.

[0113] Using the driving record data logs generated during the accident as contextual input, a large language model is used to visualize the data, generating curves and charts that change over time. The analysis of accident liability is performed by combining key memories from before, during, and after the accident, driving record descriptions, and accident snapshots, generating a structured liability report.

[0114] refer to Figure 6 and Figure 7 In this process, the large language model first generates multiple time-varying graphs based on the vehicle's driving log data, such as speed versus time, acceleration versus time, and the relative distance between the vehicle in front and behind. Furthermore, the graphs will also indicate the time or time period of the incident. Subsequently, the large language model is updated with a task—based on the time-varying graphs, combined with key memories before, during, and after the accident, driving log descriptions, and accident snapshots, to perform a causal analysis of the accident. The accident liability determination includes basic accident information, liability assessment, post-accident analysis, causal analysis, traffic violations, and the liability determination result.

[0115] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for assisting in determining liability in traffic accidents based on a large language model, characterized in that, Includes the following steps: S1. Collect data information about the vehicle's driving process before the accident occurs; S2. Integrate and transmit data information from various data sources, preprocess the data information to obtain structured data, and generate driving record descriptions; S3. Based on the structured data and driving record descriptions, call the large language model to analyze the accident situation; S31. Before the accident occurs, the large language model is called through the API interface, with structured data as input, and the response content of the model is obtained by using the prompt word template of the large language model. The model's response content is converted into a time-series format description of driving records; Based on the current time series format of the driving record description and real-time data, generate a phase summary for this time period and store it as "Key Memory - Phase Driving Record Data Summary" as the new context input for the next time series; S32. During the accident occurrence phase, the large language model determines whether a sudden event has occurred based on the driving record description, and then determines whether a traffic accident has occurred. S321. Based on the "Key Memory - Summary of Stage Driving Data Records" in S31, the large language model identifies whether there is a sudden event in the structured description; if no event occurs, the stage description of the time series is stored as "Key Memory - Driving Record Description", and the vehicle speed, acceleration, and vehicle start-up status parameters are used to determine whether the vehicle has stopped. If stopped, the large language model is invoked to summarize the trip information, generate a "key memory - summary of stage driving data records" and end the traffic accident judgment process; If the process does not stop, delete the data related to the time series from the context input of the large model, and use the staged descriptive memory as the context input for the next time series; In case of an emergency, proceed to S322; S322. In the event of an emergency, the large language model determines whether a traffic accident has occurred and generates an analysis record. In case of a traffic accident, enter S323; S323. The large language model increases the frequency of driving record descriptions and generates driving record descriptions as new contextual inputs for the large language model. S324. Large language model determines whether a traffic accident has ended; If the traffic accident is not yet over, generate a driving record description and key memory - a summary of driving record data for each stage, and store the key memory - summary of driving record data for each stage; Once the traffic accident is resolved, proceed to S33; S33. In the post-accident phase, the large language model generates descriptions and key memories with new driving record description frequencies. S34. The large language model generates an accident snapshot based on key memories described in different driving records at three time stages: before, during, and after the accident. S4. The large language model generates a structured liability assessment report based on the analysis of the vehicle's accident situation, and transmits the structured liability assessment report to the traffic management department to assist in the determination of liability for traffic accidents.

2. The traffic accident liability determination method based on a large language model according to claim 1, characterized in that, The data information in S1 includes: Vehicle information includes basic vehicle operation information, vehicle operating status information, and safety device status information; Road and environmental data, including road conditions, traffic signals, and weather data; External vehicle and environmental information, including the driving behavior of other vehicles and surrounding obstacles.

3. The traffic accident liability determination method based on a large language model according to claim 1, characterized in that, The preprocessing of data information in S2 includes: Remove noise from sensor data, fill in missing data, and correct abnormal data; Convert all data into a unified coordinate system and timestamp; External data such as weather and road conditions are transformed into descriptions with contextual semantics through feature extraction; The preprocessed structured data is transmitted to the large language model in real time and a driving record description is generated.

4. The traffic accident liability determination method based on a large language model according to claim 1, characterized in that, S4 specifically includes: Using the driving record data logs generated during the accident as contextual input, a large language model is used to visualize the data, generating curves and charts that change over time. The analysis of accident liability is performed by combining key memories from before, during, and after the accident, driving record descriptions, and accident snapshots, generating a structured liability report.

5. The traffic accident liability determination method based on a large language model according to claim 4, characterized in that, The time-varying curves include: the curve of speed changing over time, the curve of acceleration changing over time, and the curve of the relative distance between the vehicle in front and the vehicle behind.

6. The traffic accident liability determination method based on a large language model according to claim 1, characterized in that, The S322 further includes: If no traffic accident occurs, the characteristics of the emergency are described and analyzed and stored as key memories, and it is determined whether the vehicle has stopped. If the vehicle has stopped, the large language model summarizes the trip information based on all the stored key memories and ends the traffic accident judgment process. If the vehicle does not stop, during the generation of the driving record description for the next time series, the data input from the previous time series is deleted, and the key memory—the description and analysis record of the sudden event and the key memory—the summary of the driving record data in stages are used as the new context input for the next time series.

7. The traffic accident liability determination method based on a large language model according to claim 1, characterized in that, S34 specifically includes: The stored key memories, including the feature descriptions and analysis records of the sudden event and the summary of the driving record data of the previous stage before the accident, are used as new contextual inputs to the large language model to generate an accident snapshot, so as to provide an overview of the entire accident process.

Images