An all-scene traffic accident auxiliary responsibility determination intelligent agent system
Patent Information
- Application Number
- CN202610808569.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-05
- Publication Date
- 2026-08-28
AI Technical Summary
[0004]资深民警的宝贵定责经验难以系统化沉淀与传承，新民警需要经过长期实践才能独立胜任定责工作，人才培养周期长，难以适应事故处理业务量快速增长的需求
本发明通过将多模态大模型与交通法规知识图谱深度融合，构建了面向交通事故全场景的辅助定责智能体系统，实现了从“人找法”到“法找人”的智能化跨越，大幅缩短了事故处理时长。其独创的事中偏差预警机制，将传统的事后监督前移至定责过程中，有效降低了执法随意性与行政复议率。同时，系统采用松耦合的微服务架构与统一的数据治理标准，解决了法规查阅耗时、定责标准不一、监管手段滞后三大行业痛点，推动了交警事故处理模式从“经验驱动”向“数据智能驱动”的根本性转型，具备显著的行业示范意义与推广价值。
Smart Images
Figure CN122656809A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of traffic accident determination technology, specifically to an intelligent system for assisting in determining liability in all scenarios of traffic accidents. Background Technology
[0002] The current model for determining liability in traffic accidents heavily relies on the individual law enforcement experience and professional knowledge of frontline police officers. Officers are required to manually collect accident evidence, analyze case details, review relevant laws and regulations, refer to historical precedents, and ultimately determine liability during on-site or remote mediation. They also need to manually prepare various legal documents. In practice, this model has gradually revealed many insurmountable limitations: Due to differences in the professional level and law enforcement experience of different police officers, there may be significant deviations in the determination of responsibility for the same type of traffic accident. Especially in complex accident scenarios, disputes over the division of responsibility occur frequently, resulting in a high rate of administrative review and administrative litigation, which seriously affects the credibility of law enforcement.
[0003] Police officers spend a significant amount of time reviewing scattered legal provisions and historical cases during the liability determination process. Furthermore, the creation of legal documents requires the manual input of a large amount of repetitive information, resulting in lengthy processing times for individual accidents. During peak accident periods, a large backlog of accidents can easily lead to traffic congestion and public dissatisfaction.
[0004] The valuable experience of senior police officers in determining liability is difficult to systematically accumulate and pass on, and new police officers need long-term practice to be able to independently handle the work of determining liability. The talent training cycle is long and it is difficult to adapt to the rapidly increasing demand for accident handling. Summary of the Invention
[0005] The purpose of this invention is to provide an intelligent agent system for assisting in determining liability in all scenarios of traffic accidents, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, an intelligent agent system for assisting in determining liability in all scenarios of traffic accidents is provided, including an infrastructure service layer, a platform service layer, a data service layer, an application service layer, and a safety system that runs through the entire stack; The infrastructure service layer is used to provide pooled computing, storage, network and security underlying resources, and supports the unified scheduling of data center IDC national cryptographic upgrades, domestic general computing resources and domestic supercomputing resources; The platform service layer is deployed on top of the infrastructure service layer and is used to provide support for domestic basic software, scheduling of streaming computing and machine learning computing power, storage of relational / object / vector multi-databases, and common platform services including message services, GIS services, and orchestration engines. The data service layer, deployed above the platform service layer, serves as the core intelligent engine of the system. It includes a model foundation, a model training module, a data engineering module, a vertical domain model module, and a model application module. The model foundation enables cross-model collaborative calls between the Deepseek model and the Qwen multimodal model via the Model Context Protocol (MCP). The model training module covers the entire lifecycle management of model development, training, testing, evaluation, and deployment. The data engineering module is responsible for the access, processing, annotation, and task scheduling of traffic accident-related data. The vertical domain model module is used to complete the sorting of traffic accident data elements, rule sorting, legal and regulatory text parsing, major violation weight determination, and extraction of tens of thousands of historical cases, constructing a traffic law knowledge graph and vector database. The model application module integrates AI agents, AI question answering, RAG retrieval enhancement generation, and tool libraries, encapsulating business-oriented intelligent service interfaces. The application service layer is deployed on top of the data service layer and includes the accident analysis platform liability determination assistant, the PC-based liability determination assistant, and the convenient mobile liability determination assistant. The three share the intelligent kernel of the data service layer and provide differentiated intelligent assistance functions for back-end administrators, PC-based operators, and front-line police officers, respectively. The full-stack security system is independently embedded in each of the above layers, constructing a seven-dimensional protection network covering network security, host security, password security, transmission security, storage security, switching security, application security, and security assessment.
[0007] Furthermore, the traffic regulations knowledge graph and vector database constructed by the vertical domain model module transform unstructured legal provisions and historical cases into computable semantic assets, achieving accurate mapping between illegal acts and corresponding legal provisions.
[0008] Furthermore, the AI agent of the model application module integrates ASR speech translation capabilities to achieve real-time translation of remote audio and video and automatic extraction of key accident elements. Combined with the built-in rule engine, it performs logical reasoning and outputs recommendations for illegal behaviors, regulatory matching results, and preliminary suggestions for liability determination.
[0009] Furthermore, the convenient mobile liability determination assistant embeds an interactive question-and-answer robot and an anomaly detection and deviation warning module; the anomaly detection and deviation warning module has a built-in dedicated algorithm for real-time interception and review of cases with missing descriptions, incorrect legal citations, and significant deviations between manual liability determination results and system suggestions.
[0010] Furthermore, the PC-based liability determination assistant includes a smart auxiliary module for video mediation sidebar, which dynamically pushes matching traffic accident precedents and corresponding legal provisions based on the accident elements extracted in real time during remote video mediation.
[0011] Furthermore, the accident analysis platform's liability determination assistant includes a rule and case lifecycle visualization management module, which is used to realize dynamic updates of the traffic law knowledge base, visualized configuration of liability determination rules, and global situational awareness of traffic accidents.
[0012] Furthermore, the data service layer constructs an end-to-end automated pipeline to achieve fully automated processing from raw accident data collection, knowledge extraction, model iterative training to intelligent liability determination suggestions.
[0013] Furthermore, the infrastructure service layer enables dynamic allocation and elastic scaling of computing, storage, network, and security resources through a unified resource scheduling service, automatically adjusting resource quotas according to the needs of different business scenarios such as model training and real-time inference.
[0014] Furthermore, the responsibility determination process of the AI intelligent agent includes: intelligently aggregating and analyzing multimodal evidence such as accident scene photos, videos, and transcripts to construct a panoramic evidence chain; combining automatically extracted key elements such as accident time, location, collision parts, and illegal acts; and generating responsibility determination suggestions and corresponding credibility scores through data analysis, rule matching, precedent reference, and AI reasoning; and generating intelligent transcripts and traffic accident liability certificates that comply with legal norms with one click based on the generated responsibility determination suggestions and the extracted accident elements.
[0015] Furthermore, the complete operation process of the system includes the following steps: S1. The data engineering module continuously accesses the original data from the traffic violation database, the database of tens of thousands of historical accident cases, and the database of current traffic laws and regulations. After completing data cleaning, structured labeling, and permission classification, the data is pushed to the vertical domain model module. S2, the vertical domain model module extracts elements and performs semantic transformation on the input data, constructs and dynamically updates the traffic law knowledge graph and vector database, and trains and generates a weight judgment model for major violations. S3. The application service layer receives accident handling requests from different terminals, obtains multimodal evidence data including accident scene photos, videos, voice transcripts, and statements from the parties involved, and transmits them to the data service layer. S4. The AI agent in the model application module calls the Qwen multimodal model and ASR capability to automatically extract key elements for determining liability, such as accident time, location, collision part, and illegal behavior, to construct a complete chain of evidence and calculate the reliability of the evidence. S5 and AI intelligent agents use RAG search enhancement generation technology to match corresponding legal provisions and similar historical precedents, and combine rule engine and major illegal behavior weight judgment model to perform multi-dimensional logical reasoning to generate liability determination suggestions that include liability division results, legal basis and credibility score; S6. The system will push the responsibility determination suggestions to the corresponding terminal processing personnel. At the same time, the anomaly detection and deviation warning module of the convenient mobile terminal will monitor the manual responsibility determination process in real time, and intercept and trigger the review process in real time for missing case descriptions, incorrect legal citations and major deviations in responsibility determination results. S7. After the personnel handling the case confirm or adjust the liability determination results, the AI agent automatically extracts all accident elements and liability determination basis, and generates intelligent records and traffic accident determination letters that comply with legal norms with one click. S8. The complete handling data, liability determination results, and documents of this accident are automatically transmitted back to the data engineering module for iterative optimization of the model and dynamic updating of the knowledge base, forming a full-link intelligent closed loop of data, model, application, and feedback.
[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention deeply integrates a multimodal large-scale model with a traffic regulation knowledge graph to construct an intelligent system for assisting in determining liability across all traffic accident scenarios. This achieves an intelligent leap from "people searching for the law" to "the law finding people," significantly shortening accident processing time. Its unique in-process deviation early warning mechanism moves traditional post-event supervision forward to the liability determination process, effectively reducing arbitrariness in enforcement and the rate of administrative review. Simultaneously, the system adopts a loosely coupled microservice architecture and unified data governance standards, addressing three major industry pain points: time-consuming regulation searches, inconsistent liability determination standards, and lagging regulatory methods. This drives a fundamental transformation in traffic police accident handling from "experience-driven" to "data-driven intelligence," demonstrating significant industry demonstration value and potential for widespread adoption. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the architecture of the auxiliary responsibility determination assistant system of the present invention; Figure 2 This is a schematic diagram of the operation interface of the accident liability determination assistant system of the present invention. Detailed Implementation
[0018] Please see Figures 1-2 This invention provides an intelligent agent system for assisting in determining liability in all scenarios of traffic accidents, including an infrastructure service layer, a platform service layer, a data service layer, an application service layer, and a security system that runs through the entire stack; The infrastructure service layer is used to provide pooled computing, storage, networking and security underlying resources, and supports the unified scheduling of data center IDC national cryptographic upgrades, domestic general computing resources and domestic supercomputing resources; The platform service layer is deployed on top of the infrastructure service layer to provide support for domestic basic software, scheduling of streaming computing and machine learning computing power, storage of relational / object / vector multi-databases, and common platform services including message services, GIS services, and orchestration engines; The data service layer, deployed above the platform service layer, serves as the core intelligent engine of the system. It includes a model foundation, a model training module, a data engineering module, a vertical domain model module, and a model application module. The model foundation enables cross-model collaborative calls between the Deepseek model and the Qwen multimodal model via the Model Context Protocol (MCP). The model training module covers the entire lifecycle management of model development, training, testing, evaluation, and deployment. The data engineering module is responsible for the access, processing, annotation, and task scheduling of traffic accident-related data. The vertical domain model module is used to complete the sorting of traffic accident data elements, rule sorting, legal and regulatory text parsing, major violation weight determination, and extraction of tens of thousands of historical cases, constructing a traffic law knowledge graph and vector database. The model application module integrates AI agents, AI question answering, RAG retrieval enhancement generation, and tool libraries, encapsulating business-oriented intelligent service interfaces. The application service layer is deployed on top of the data service layer and includes the accident analysis platform liability determination assistant, the PC-based liability determination assistant, and the convenient mobile liability determination assistant. The three share the intelligent core of the data service layer and provide differentiated intelligent assistance functions for back-end administrators, PC-based operators, and front-line police officers, respectively. The full-stack security system is independently embedded in each of the above layers, constructing a seven-dimensional protection network covering network security, host security, password security, transmission security, storage security, switching security, application security, and security assessment.
[0019] The traffic regulations knowledge graph and vector database constructed by the vertical domain model module transform unstructured legal provisions and historical cases into computable semantic assets, enabling accurate mapping between illegal acts and corresponding legal provisions.
[0020] The AI agent in the model application module integrates ASR speech translation capabilities, enabling real-time translation of remote audio and video and automatic extraction of key accident elements. Combined with the built-in rule engine, it performs logical reasoning and outputs recommendations for illegal behaviors, regulatory matching results, and preliminary suggestions for liability determination.
[0021] The Convenient Mobile Liability Determination Assistant embeds an interactive Q&A robot, as well as an anomaly detection and deviation warning module. The anomaly detection and deviation warning module has a built-in proprietary algorithm for real-time interception and review of cases with missing descriptions, incorrect legal citations, and significant deviations between manual liability determination results and system recommendations.
[0022] The PC-based liability determination assistant includes a smart assistance module in the video mediation sidebar, which dynamically pushes matching traffic accident precedents and corresponding legal provisions based on real-time extracted accident elements during remote video mediation.
[0023] The accident analysis platform's liability determination assistant includes a visualization management module for the entire lifecycle of rules and cases, which enables dynamic updates of the traffic law knowledge base, visualization configuration of liability determination rules, and global situational awareness of traffic accidents.
[0024] The data service layer constructs an end-to-end automated pipeline to achieve fully automated processing from raw accident data collection, knowledge extraction, model iterative training to intelligent liability determination suggestions.
[0025] The infrastructure service layer enables dynamic allocation and elastic scaling of computing, storage, network, and security resources through a unified resource scheduling service, automatically adjusting resource quotas according to the needs of different business scenarios such as model training and real-time inference.
[0026] The AI-powered liability determination process includes: intelligently aggregating and analyzing multimodal evidence such as accident scene photos, videos, and transcripts to construct a panoramic evidence chain; combining automatically extracted key elements such as accident time, location, collision points, and illegal acts; and generating liability determination suggestions and corresponding credibility scores through data analysis, rule matching, precedent reference, and AI reasoning. Based on the generated liability determination suggestions and all extracted accident elements, a legally compliant intelligent transcript and traffic accident liability certificate are generated with one click.
[0027] The complete operation process of the system includes the following steps: S1. The data engineering module continuously accesses the original data from the traffic violation database, the database of tens of thousands of historical accident cases, and the database of current traffic laws and regulations. After completing data cleaning, structured labeling, and permission classification, the data is pushed to the vertical domain model module. S2, the vertical domain model module extracts elements and performs semantic transformation on the input data, constructs and dynamically updates the traffic law knowledge graph and vector database, and trains and generates a weight judgment model for major violations. S3. The application service layer receives accident handling requests from different terminals, obtains multimodal evidence data including accident scene photos, videos, voice transcripts, and statements from the parties involved, and transmits them to the data service layer. S4. The AI agent in the model application module calls the Qwen multimodal model and ASR capability to automatically extract key elements for determining liability, such as accident time, location, collision part, and illegal behavior, to construct a complete chain of evidence and calculate the reliability of the evidence. S5 and AI intelligent agents use RAG search enhancement generation technology to match corresponding legal provisions and similar historical precedents, and combine rule engine and major illegal behavior weight judgment model to perform multi-dimensional logical reasoning to generate liability determination suggestions that include liability division results, legal basis and credibility score; S6. The system will push the responsibility determination suggestions to the corresponding terminal processing personnel. At the same time, the anomaly detection and deviation warning module of the convenient mobile terminal will monitor the manual responsibility determination process in real time, and intercept and trigger the review process in real time for missing case descriptions, incorrect legal citations and major deviations in responsibility determination results. S7. After the personnel handling the case confirm or adjust the liability determination results, the AI agent automatically extracts all accident elements and liability determination basis, and generates intelligent records and traffic accident determination letters that comply with legal norms with one click. S8. The complete handling data, liability determination results, and documents of this accident are automatically transmitted back to the data engineering module for iterative optimization of the model and dynamic updating of the knowledge base, forming a full-link intelligent closed loop of data, model, application, and feedback.
Claims
1. A smart agent system for assisting in determining liability across all scenarios of traffic accidents, characterized in that, This includes the infrastructure service layer, platform service layer, data service layer, application service layer, and a security system that runs throughout the entire stack; The infrastructure service layer is used to provide pooled computing, storage, network and security underlying resources, and supports the unified scheduling of data center IDC national cryptographic upgrades, domestic general computing resources and domestic supercomputing resources; The platform service layer is deployed on top of the infrastructure service layer and is used to provide support for domestic basic software, scheduling of streaming computing and machine learning computing power, storage of relational / object / vector multi-databases, and common platform services including message services, GIS services, and orchestration engines. The data service layer, deployed above the platform service layer, serves as the core intelligent engine of the system. It includes a model foundation, a model training module, a data engineering module, a vertical domain model module, and a model application module. The model foundation enables cross-model collaborative calls between the Deepseek model and the Qwen multimodal model via the Model Context Protocol (MCP). The model training module covers the entire lifecycle management of model development, training, testing, evaluation, and deployment. The data engineering module is responsible for the access, processing, annotation, and task scheduling of traffic accident-related data. The vertical domain model module is used to complete the sorting of traffic accident data elements, rule sorting, legal and regulatory text parsing, major violation weight determination, and extraction of tens of thousands of historical cases, constructing a traffic law knowledge graph and vector database. The model application module integrates AI agents, AI question answering, RAG retrieval enhancement generation, and tool libraries, encapsulating business-oriented intelligent service interfaces. The application service layer is deployed on top of the data service layer and includes the accident analysis platform liability determination assistant, the PC-based liability determination assistant, and the convenient mobile liability determination assistant. The three share the intelligent kernel of the data service layer and provide differentiated intelligent assistance functions for back-end administrators, PC-based operators, and front-line police officers, respectively. The full-stack security system is independently embedded in each of the above layers, constructing a seven-dimensional protection network covering network security, host security, password security, transmission security, storage security, switching security, application security, and security assessment.
2. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The traffic regulations knowledge graph and vector database constructed by the vertical domain model module transform unstructured legal provisions and historical cases into computable semantic assets, enabling accurate mapping between illegal acts and corresponding legal provisions.
3. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The AI agent in the model application module integrates ASR speech translation capabilities, enabling real-time translation of remote audio and video and automatic extraction of key accident elements. Combined with the built-in rule engine, it performs logical reasoning and outputs recommendations for illegal behaviors, regulatory matching results, and preliminary suggestions for liability determination.
4. The assisted responsibility-determination intelligent agent system according to claim 1, characterized in that, The convenient mobile liability determination assistant embeds an interactive question-and-answer robot and an anomaly detection and deviation warning module; the anomaly detection and deviation warning module has a built-in dedicated algorithm for real-time interception and review of cases with missing descriptions, incorrect legal citations, and significant deviations between manual liability determination results and system suggestions.
5. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The PC-based liability determination assistant includes a smart auxiliary module in the video mediation sidebar, which dynamically pushes matching traffic accident precedents and corresponding legal provisions based on real-time extracted accident elements during remote video mediation.
6. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The accident analysis platform's liability determination assistant includes a rules and case lifecycle visualization management module, which is used to realize dynamic updates of the traffic law knowledge base, visualized configuration of liability determination rules, and global situational awareness of traffic accidents.
7. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The data service layer constructs an end-to-end automated pipeline to achieve fully automated processing from raw accident data collection, knowledge extraction, model iterative training to intelligent liability determination suggestions.
8. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The infrastructure service layer achieves dynamic allocation and elastic scaling of computing, storage, network and security resources through a unified resource scheduling service, and automatically adjusts resource quotas according to the needs of different business scenarios such as model training and real-time inference.
9. The auxiliary responsibility-determination intelligent agent system according to claim 3, characterized in that, The AI agent's liability determination process includes: intelligently aggregating and analyzing multimodal evidence such as accident scene photos, videos, and transcripts to construct a panoramic evidence chain; combining automatically extracted key elements such as accident time, location, collision points, and illegal acts; and generating liability determination suggestions and corresponding credibility scores through data analysis, rule matching, precedent reference, and AI reasoning. Based on the generated liability determination suggestions and all extracted accident elements, a legally compliant intelligent transcript and traffic accident liability certificate are generated with one click.
10. The auxiliary responsibility-determination intelligent agent system according to claim 1, characterized in that, The complete operation process of the system includes the following steps: S1. The data engineering module continuously accesses the original data from the traffic violation database, the database of tens of thousands of historical accident cases, and the database of current traffic laws and regulations. After completing data cleaning, structured labeling, and permission classification, the data is pushed to the vertical domain model module. S2, the vertical domain model module extracts elements and performs semantic transformation on the input data, constructs and dynamically updates the traffic law knowledge graph and vector database, and trains and generates a weight judgment model for major violations. S3. The application service layer receives accident handling requests from different terminals, obtains multimodal evidence data including accident scene photos, videos, voice transcripts, and statements from the parties involved, and transmits them to the data service layer. S4. The AI agent in the model application module calls the Qwen multimodal model and ASR capability to automatically extract key elements for determining liability, such as accident time, location, collision part, and illegal behavior, to construct a complete chain of evidence and calculate the reliability of the evidence. S5 and AI intelligent agents use RAG search enhancement generation technology to match corresponding legal provisions and similar historical precedents, and combine rule engine and major illegal behavior weight judgment model to perform multi-dimensional logical reasoning to generate liability determination suggestions that include liability division results, legal basis and credibility score; S6. The system will push the responsibility determination suggestions to the corresponding terminal processing personnel. At the same time, the anomaly detection and deviation warning module of the convenient mobile terminal will monitor the manual responsibility determination process in real time, and intercept and trigger the review process in real time for missing case descriptions, incorrect legal citations and major deviations in responsibility determination results. S7. After the personnel handling the case confirm or adjust the liability determination results, the AI agent automatically extracts all accident elements and liability determination basis, and generates intelligent records and traffic accident determination letters that comply with legal norms with one click. S8. The complete handling data, liability determination results, and documents of this accident are automatically transmitted back to the data engineering module for iterative optimization of the model and dynamic updating of the knowledge base, forming a full-link intelligent closed loop of data, model, application, and feedback.