A deep learning-based road responsibility automatic determination system

Abstract

The application discloses a kind of road responsibility automatic determination systems based on deep learning, belong to intelligent transportation technical field.System includes: multi-source data cooperation and evidence collection module, for obtaining and solidifying the field data containing space-time information, forms tamper-evident evidence chain;Fusion deep learning and rule reasoning determination engine, for parsing evidence chain, through perception network extraction target and event characteristics, in the unified space-time scene fusion road network and ownership information, and based on configurable rule knowledge base carries out logical reasoning and quantitative determination;Lightweight road right digital twin and visualization module, for the determination result and three-dimensional scene fusion generation interactive report.The application realizes the whole process automation from data credible collection, intelligent analysis to conclusion visualization, solves the problem that traditional method is low in efficiency, strong in subjectivity, process is not transparent, can be widely applied to traffic accident responsibility, facility maintenance responsibility division and road right management.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation, computer vision and data processing technology, and in particular to an automatic road liability determination system based on deep learning. Background Technology

[0002] Currently, road liability determination (such as determining liability in traffic accidents and allocating liability for damage caused by defects in road facilities) mainly relies on manual on-site investigations, playback of surveillance videos, statements from the parties involved, and experience-based judgment. This approach has several drawbacks: First, it is inefficient and time-consuming, especially when there is abundant evidence or complex scenarios; second, it is highly subjective, as different individuals' experiences and judgment standards can lead to differing conclusions; third, the process lacks transparency, making it difficult for the responsible party to fully understand the basis for the decision; and finally, with the development of vehicle-road cooperation and high-precision maps, road management is becoming increasingly sophisticated, and traditional methods struggle to process massive amounts of multi-source, heterogeneous data and accurately correlate it with complex legal provisions.

[0003] While existing technologies include accident detection based on dashcams or liability analysis using video tracking, they largely focus on single data sources or post-event video analysis, lacking a complete closed loop from reliable multi-source data collection to automatic rule matching. Other technologies involve traffic situation prediction and simulation, but these are not deeply integrated with specific liability determination rules. Summary of the Invention

[0004] This invention aims to overcome the shortcomings of existing technologies and solve technical problems such as high reliance on human intervention, low efficiency, strong subjectivity, disconnect between evidence and rules, and unintuitive presentation of conclusions in the process of determining road liability.

[0005] To achieve the above objectives, the basic solution of the present invention is as follows: To achieve the above objectives, this invention proposes an automatic road responsibility determination system based on deep learning.

[0006] The system includes: The multi-source data collaborative acquisition and evidence collection module is used to synchronously collect multimodal data from the event scene and perform spatiotemporal alignment, and generate a tamper-proof evidence chain based on chain hashing and trusted timestamp technology; A decision engine integrating deep learning and rule-based reasoning, connected to the evidence collection module, is used to receive and parse the evidence chain, and includes: A multimodal perception and understanding unit is used to identify traffic participants, road facilities and environmental elements from the multimodal data, and output their status and spatiotemporal trajectory; The spatiotemporal scene reconstruction and reasoning unit is used to construct a unified spatiotemporal scene model that integrates dynamic trajectories, static road network information and ownership rules, and to perform event inference and quantitative analysis based on the model. A configurable rule knowledge base stores computer-executable logical rules that transform liability determination regulations into their equivalents. The responsibility reasoning and judgment unit is used to extract events from the spatiotemporal scene model, call the rule knowledge base for matching and reasoning, and output responsibility judgment conclusions and basis. The lightweight right-of-way digital twin and visualization module, connected to the determination engine, is used to integrate the responsibility determination conclusion and process data with the three-dimensional real-scene model to generate an interactive visualization report.

[0007] Preferably, the multi-source data collaborative acquisition and evidence collection module includes an integrated acquisition terminal deployed on a vehicle platform, drone, or roadside facility. The terminal has a built-in unified timing unit for aligning video frames, point cloud data, inertial measurement unit (IMU) data, and global navigation satellite system (GNSS) data within time windows. The module generates content hashes for each window of data in chronological order and concatenates the current content hash with the previous hash chain value to calculate a new chain value, forming a chain structure. The terminal security key is used to digitally sign the record containing the chain value and bind a timestamp from a trusted third party, and finally writes it to a write-once, read-many (WORM) memory.

[0008] Preferably, the multimodal perception and understanding unit includes a multi-task neural network based on a shared feature extraction backbone, which performs the following tasks in parallel: detecting, segmenting, and tracking traffic participants across frames; performing semantic segmentation and state recognition on lane lines, traffic signs, and traffic lights; and detecting and assessing the severity of road surface defects.

[0009] Preferably, the spatiotemporal scene reconstruction and reasoning unit performs the following operations: based on sensor calibration parameters, it maps the target trajectory and facility location output by the multimodal perception and understanding unit to a unified world coordinate system; it imports a high-precision digital map containing lane topology, right-of-way boundaries, and facility ownership information; it constructs a spatiotemporal scene model using a graph structure, where nodes represent physical or logical entities, and edges represent spatiotemporal, ownership, or rule relationships between entities; the unit also integrates a traffic flow simulation model, used to simulate the event occurrence process based on the spatiotemporal scene model, and output collision dynamics parameters and avoidability analysis indicators.

[0010] Preferably, the rules in the configurable rule knowledge base are represented in the form of "condition-conclusion". The conditions are associated with node attributes, edge relationships or event types in the spatiotemporal scene model, and the conclusions include the responsible entity identifier, fault type and weight coefficient. The knowledge base provides a graphical management interface that supports the editing, retrieval and priority configuration of rules.

[0011] Preferably, the workflow of the responsibility reasoning and judgment unit includes: Event extraction steps: Based on preset spatiotemporal and state thresholds, automatically identify potential violations or facility defects related events from the spatiotemporal scene model; Rule matching and triggering steps: Match the extracted event features with the conditions in the rule knowledge base to activate all applicable rules; Responsibility quantification step: Calculate the responsibility contribution of each relevant party by combining the weight coefficients of the triggered rules and the quantitative analysis indicators output by the spatiotemporal scene reconstruction and reasoning unit; Report generation steps: Integrate the results of responsibility allocation, the list of triggered rules, the index of key evidence, and a summary of the analysis process to form a structured judgment report.

[0012] Preferably, the lightweight right-of-way digital twin and visualization module is implemented through a browser-accessible Web 3D engine; the 3D real-world model it loads is constructed from oblique photogrammetry point cloud data and optimized by level of detail (LOD); the module supports highlighting responsible parties in the 3D scene, dynamically replaying event processes, interactively querying the attributes and ownership information of any facility, and allowing users to adjust parameters to perform responsibility simulation.

[0013] The system constructs a complete technology chain that integrates "end-edge-cloud": at the data source, chained hashing and trusted timestamps ensure the authenticity and integrity of the original data; at the analysis center, multi-task deep learning models are used to achieve accurate perception and understanding of complex scenarios, and a unified "spatiotemporal-ownership" scene graph is constructed to integrate the dynamic and static rules of the physical world, and an automated reasoning is performed with the help of a configurable rule engine; at the result presentation end, a lightweight 3D engine is used to project the abstract judgment process and conclusions into the real-world digital twin model, providing immersive and interactive visualization reports.

[0014] This invention also provides an automatic road liability determination method based on this system, comprising the following steps: S1: Simultaneously acquire on-site data through the multi-source data collaborative acquisition and evidence collection module, and generate a tamper-proof evidence chain; S2: The evidence chain is analyzed through the judgment engine that integrates deep learning and rule reasoning to complete environmental perception, scene reconstruction and event analysis; S3: In the reconstructed spatiotemporal scene model, automated logical reasoning and responsibility quantification are performed based on a configurable rule knowledge base; S4: The judgment results are fused and presented interactively through the lightweight right-of-way digital twin and visualization module.

[0015] Compared with existing technologies, the road responsibility automatic determination system based on deep learning proposed in this application has the following advantages: 1. Full-process automation and efficiency improvement: The entire process from on-site data collection and analysis to report generation is automated, which greatly reduces manual intervention and significantly shortens the responsibility determination cycle.

[0016] 2. Credible Evidence and Traceable Process: Based on cryptographic principles, a tamper-proof evidence chain is constructed to ensure the authenticity of the basic data for analysis. Furthermore, all key steps in the entire judgment process are traceable and auditable, enhancing the legal validity of the conclusion.

[0017] 3. Objective and accurate judgment: Features are extracted uniformly through deep learning models, and logical judgments are made based on a clear and configurable rule base, which minimizes the influence of subjective factors and improves the accuracy of responsibility division in complex scenarios.

[0018] 4. Flexible and scalable rules: The configurable rule knowledge base allows administrators to flexibly add, delete, or modify rules according to local regulations or management needs, giving the system good adaptability and scalability. Application scenarios can be extended from traffic accidents to multiple fields such as road facility operation and maintenance and right-of-way management.

[0019] 5. The conclusions are intuitive and highly interactive: The use of lightweight digital twin technology to generate three-dimensional visualization reports makes the spatiotemporal relationships, behavioral processes, and basis for responsibility clear at a glance, facilitating understanding, communication, and verification by all parties, and enhancing public trust and law enforcement transparency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the automatic road responsibility determination system provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the workflow of the multi-source data collaborative acquisition and evidence collection module in this embodiment of the invention.

[0022] Figure 3 This is a flowchart illustrating the internal structure and data processing of the decision engine that integrates deep learning and rule-based reasoning in an embodiment of the present invention.

[0023] Figure labeling: 100: Multi-source data collaborative acquisition and evidence collection module; 200: Judgment engine integrating deep learning and rule reasoning; 300: Lightweight right-of-way digital twin and visualization module; 210: Multimodal perception and understanding unit; 220: Spatiotemporal scene reconstruction and reasoning unit; 230: Configurable rule knowledge base; 240: Responsibility reasoning and judgment unit. Detailed Implementation

[0024] The following detailed description illustrates the specific implementation method: like Figure 1As shown, this embodiment provides an automatic road responsibility determination system based on deep learning. Its core lies in constructing an intelligent judgment closed loop that is data-driven, rule-oriented, and provides visual feedback. The system mainly includes three functional modules: I. Multi-source data collaborative acquisition and evidence collection module (100) like Figure 2 As shown, this module serves as the "trust anchor" for system data input. When an event requiring liability determination occurs (such as a traffic accident or the discovery of serious road defects), mobile platforms equipped with multiple sensors (such as reconnaissance drones or police patrol vehicles) or fixed roadside equipment can be activated for data collection. The data acquisition terminal has a built-in high-precision clock source to ensure strict time synchronization of sensor data from video cameras, LiDAR, inertial measurement units (IMUs), and Global Navigation Satellite System (GNSS) receivers.

[0025] The key innovation lies in the construction of the evidence chain: the system calculates the content hash value of the raw or feature data of each sensor within a fixed time window (e.g., 33ms). Then, the device's unique identifier, precise timestamp, the content hash value of the current window, and the chained hash value calculated in the previous time window are concatenated in a fixed order and hashed again to obtain a new chained hash value for the current window. This process ensures that data blocks are linked sequentially, and any tampering with historical data will invalidate all subsequent chain values. Next, a key stored in the terminal's secure hardware is used to digitally sign the complete record (including the original data index, timestamp, chain value, etc.). Simultaneously, the module applies for an authoritative timestamp from the National Time Service Center or a Trusted Time Stamp Service (TSA) via the network and binds it to the signed record. Finally, these signed and timestamp-bound records are sequentially written to a storage medium with "Write Once, Read Many Times" (WORM) characteristics, forming an undeniable, immutable, and verifiable electronic evidence chain. This lays a solid legal foundation for the credibility of all subsequent analytical conclusions.

[0026] II. A decision engine that integrates deep learning and rule-based reasoning (200) like Figure 3 As shown, this module is the system's "intelligent analysis hub." It receives trusted chain of evidence data from module (100) and performs the following core steps: 1. Multimodal Perception and Understanding Unit (210): This unit integrates a multi-task deep learning model trained on a large amount of labeled data. The model development and training are as follows: The model adopts a shared feature extraction network with ResNet-101 or Swin Transformer as the backbone, and connects multiple task-specific head networks in parallel on this basis. The model training uses a combination of multiple public datasets (such as BDD100K, Cityscapes, KITTI) and a self-built dataset containing rich annotations of Chinese road scenes. The training data annotations include: bounding boxes, instance segmentation masks and cross-frame IDs of vehicles, pedestrians and non-motorized vehicles; semantic segmentation and status labels of lane lines, traffic signs and traffic lights; and annotation boxes and severity levels of road potholes and cracks. The training uses an end-to-end multi-task loss function for optimization, with the total loss L_total = αL_det + βL_seg + γL_track + δL_attr, where each item corresponds to the detection, segmentation, tracking and attribute classification loss, respectively, and α, β, γ, δ are balancing hyperparameters. Dedicated task head: (1) Traffic element perception head: responsible for detecting and segmenting vehicles, pedestrians and non-motorized vehicles in the image, and assigning a unique ID to each instance. By fusing appearance re-identification (ReID) features and motion prediction, it achieves stable tracking across frames and viewpoints.

[0027] (2) Road environment perception head: responsible for semantic segmentation of lane lines (distinguishing between real and virtual), traffic signs, ground markings (such as arrows and stop lines), and identifying the current status of traffic lights (red, yellow, green and arrow direction).

[0028] (3) Facility status sensing head: For high-definition images or LiDAR point clouds, identify defects such as cracks, potholes, and bumps on the road surface, and classify them into severity levels according to preset standards (such as area and depth).

[0029] The unit outputs structured perception results, including target trajectory sequences with IDs, road environment semantic maps, and lists of facility defects.

[0030] 2. Spatiotemporal Scene Reconstruction and Inference Unit (220): The goal of this unit is to construct a digital event scene that a computer can "understand". First, using the calibration parameters (intrinsic and extrinsic parameters) of the sensors and depth estimation technology, the trajectory and position in the 2D image coordinates output by the unit (210) are accurately back-projected into the 3D coordinate system of the real world, usually converted into a bird's-eye view (BEV). Then, a high-precision digital map of the road segment is imported. The map not only contains lane geometry information, but also embeds right-of-way management data, such as: ownership of each lane, maintenance unit, speed limit regulations, special traffic rules (such as tidal lanes, bus-only lanes), etc.

[0031] The specific data structure definition and construction process of the spatiotemporal-ownership scenario graph The "Spatio-Temporal and Ownership SceneGraph" (STO-SG) constructed in this unit is a formal graph data structure G = (V, E, A_V, A_E). As a core model for uniformly representing multi-source heterogeneous data, it forms the basis for subsequent reasoning and judgment. Its construction process involves a technical workflow that deeply integrates multimodal perception results, high-precision maps, and ownership information.

[0032] 1. Schema Definition of Graph Structure Data Node (V): Represents an entity in the scene, divided into three categories: Dynamic entity node (V_d): such as traffic participants (vehicles, pedestrians). The attribute set A_Vd includes: Physical feature attributes: directly mapped from sensor data. Such as: type (category), instance_id (tracking ID), position(t) (time-series 3D coordinates, from sensor fusion and coordinate backprojection), velocity(t) (velocity vector), bounding_box (3D bounding box), heading(t) (heading angle).

[0033] Ownership / Status Attributes: Associations inferred from maps or rules. For example: owner (vehicle owner / company, which can be associated with an external database), current_lane_id (current lane ID, obtained by matching with a high-precision map), legal_status (current legal status, such as "driving normally" or "illegally parked", which is initially inferred from rules).

[0034] Static entity nodes (V_s): such as road facilities (lane lines, traffic signs, traffic lights, potholes). The attribute set A_Vs includes: Physical feature attributes: type, geometry (geometric shape, such as the broken line of a lane line, the polygonal region of a pothole, derived from semantic segmentation and point cloud clustering), spatial_location (world coordinates), state (such as the traffic light "red", the pothole "8cm deep").

[0035] Ownership / Management Attributes: owner (ownership unit), maintainer (maintenance unit), jurisdiction (jurisdictional authority), related_regulation_id (related regulation entry ID).

[0036] Logical entity node (V_l): such as events ("lane change", "collision"), rule instances ("rule R101 was triggered"). The attribute set A_Vl includes: event_type, start_time, end_time, and involved_entities (a list of involved V_d or V_s node IDs).

[0037] Edge (E): Represents the relationship between entities, and is divided into four categories: Spatiotemporal relationship edge (E_st): Describes the physical spatiotemporal connection between entities. For example, (Vehicle_A) --[is_on]--> (Lane_L), the attribute A_Est can contain confidence and temporal_range (the time period during which the relationship is valid).

[0038] Ownership relationship edge (E_own): Describes ownership and management responsibility. For example, (Pothole_P) --[belongs_to]--> (Company_X), the attribute A_Eown can contain legal_document_ref (ownership document index).

[0039] Participation / causal relationship edge (E_rel): Describes the logical connection between an event and an entity. For example, (Event_Collision) --[involves]--> (Vehicle_A), (Rule_R101) --[triggered_by]-->(Vehicle_A).

[0040] Hierarchical / topological relationship edges (E_top): Describe the spatial containment and road topology. For example, (Lane_L1) --[connects_to]--> (Lane_L2), (Road_R) --[contains]--> (Lane_L).

[0041] 2. The process of integrating unstructured data into STO-SG This workflow fully demonstrates the fusion processing of multi-source heterogeneous data, including video (image sequences), point clouds, map vectors, and attribute databases. Step 1: Data Alignment and Coordinate Unification. Using sensor calibration parameters, SLAM (Simultaneous Localization and Mapping), or direct georeferencing techniques, all pixels in each frame of the image and every point in the point cloud are transformed into a unified global world coordinate system (such as the UTM coordinate system). This is the foundation for achieving spatiotemporal alignment.

[0042] Step 2: Multimodal Perception and Feature Extraction Mapping. The output of the multimodal perception and understanding unit (210) is the initial source for building nodes. For video streams, 2D trajectory boxes and IDs of vehicles and pedestrians are obtained through object detection and tracking models. The 2D trajectories are upscaled to a coarse position sequence position(t) in 3D space through a depth estimation model (or combined with sparse point clouds). For laser point clouds, the 3D geometry and spatial location of facilities such as road potholes and guardrails are accurately obtained through point cloud segmentation and classification algorithms. These perception results are directly mapped to the physical feature attributes of dynamic / static entity nodes.

[0043] Step 3: Map and Ownership Information Fusion. Import a high-precision vector map (including lane-level geometry and topology) and an ownership attribute database to supplement nodes with static relationships and ownership attributes. Through map matching, determine the vehicle node's current_lane_id and dynamically create spatiotemporal relationship edges E_st. Through ownership association, perform spatial queries in the ownership spatial database to associate attributes such as owner and maintainer of the facility, and create ownership relationship edges E_own.

[0044] Step 4: Graph Structure Instantiation and Relation Inference. The system instantiates STO-SG in memory or a graph database. Based on spatiotemporal proximity, kinematic formulas, and traffic rules, the system infers and creates higher-level relation edges. For example, by analyzing the trajectories and speeds of two vehicles, it automatically creates E_st edges such as "approaching" and "following"; when a vehicle is detected entering the intersection conflict zone during a red light, the system creates a logical event node V_l ("running a red light") and a corresponding participation relation edge E_rel. This process is dynamic; as time progresses and new sensory data is input, the graph structure is continuously updated, forming a dynamically evolving "spatiotemporal-ownership" digital twin scene rich in physical features and ownership semantics.

[0045] The essence of the technical solution is as follows: As can be seen from the above specific definitions and processes, the spatiotemporal scene reconstruction and reasoning unit (220) of this invention is essentially: 1) Deep feature extraction and unified representation: transforming unstructured pixels and point clouds into structured feature attributes with clear physical meaning; 2) Cross-modal data fusion: under a unified spatiotemporal benchmark, fusing and encoding information from vision, point clouds, maps, and databases into the nodes and edges of the same graph data model (STO-SG); 3) Relationship reasoning and scene understanding: performing relationship reasoning based on spatiotemporal logic and traffic knowledge based on the fused graph structure. It is this STO-SG generated by deep fusion processing that provides a precise structured and semantic foundation for subsequent rule matching and quantitative analysis.

[0046] In addition, unit (220) integrates a traffic flow simulation model for simulating the event occurrence process based on the spatiotemporal scenario model. This embodiment uses a parameter-calibrated Intelligent Driver Model (IDM) combined with a Lane Change Model that minimizes overall braking intensity (MOBIL) to construct a microscopic traffic simulator. The simulator uses the actual trajectories of traffic participants perceived within seconds before the accident as the initial state, and performs multiple Monte Carlo simulations under the same road and environmental constraints to extrapolate the possible development of the event. The simulation outputs key quantitative analysis indicators, such as: the theoretical safe speed (v_safe), minimum necessary braking deceleration (a_brake_min), collision time (TTC), and the "avoidable percentage" (P_avoidable) of the accident after considering reaction time for each participant at critical time points in the event. These indicators provide an objective dynamic basis for subsequent responsibility quantification.

[0047] 3. Configurable Rule Knowledge Base (230): This is a core component that transforms laws, regulations, and management rules described in natural language into computer-executable logical rules. Rules are in the form of "IF (condition) THEN (conclusion)". For example: Rule R101: IF [Subject type is motor vehicle] and [Traffic light status is 'red'] and [Vehicle front axle crosses the stop line] and [Vehicle enters the intersection conflict zone] THEN [Responsible subject = the vehicle, Fault = 'Violation of traffic signal', Base weight = 0.8] Rule R205: IF [Facility type is 'Pothole'] and [Pothole area > 0.1㎡] and [Pothole depth > 5cm] and [Current time - Last maintenance time > Specified cycle] THEN [Responsible entity = Facility owner, Fault = 'Failure to fulfill timely maintenance obligations', Basic weight = 0.6] Administrators can easily maintain, update, and test the rule base through a graphical interface.

[0048] 4. Specific workflow and quantitative model of the responsibility reasoning and judgment unit (240): This unit is the final decision generator, and its workflow includes the following specific steps: Event extraction steps: Based on preset spatiotemporal and state thresholds, the system automatically identifies potential violations or facility defects related events from the spatiotemporal scene model. For example, the system will detect whether a lane change occurs within a solid line area, or whether the vehicle speed continuously exceeds the speed limit of the current road segment.

[0049] Rule matching and triggering steps: Match the extracted event features with the conditions in the rule knowledge base to activate all applicable rules (the method is as described in the previous section). Each triggered rule Ri outputs a preliminary conclusion, including the responsible party, the type of fault, and a basic responsibility weight W_i_base (usually between 0 and 1, preset by domain experts when defining the rule, representing the relative severity of the fault in typical circumstances).

[0050] Responsibility quantification steps: This is the core of the invention. The system integrates the weights of all triggered rules and the quantitative analysis indicators output by the simulation unit (220), and calculates the final responsibility ratio of each relevant party through a responsibility contribution aggregation model. This embodiment adopts the following calculation model: Calculate the initial responsibility score S_k for each subject: For each responsible subject (party involved) k, its initial responsibility score is the sum of the base weights of all triggered rules pointing to it, i.e., S_k_initial = Σ_{i, subject(k)∈Ri} W_i_base.

[0051] Simulation correction factors are applied: Quantitative indicators obtained from the simulation unit (such as speeding percentage O_k, reaction delay D_k, and avoidability percentage P_avoidable_k) are used to dynamically correct the initial score. This embodiment defines a correction function F. For example, for rules triggered by behavioral errors (such as speeding), the corresponding weight will be amplified by the speeding percentage, and the corrected principal responsibility score S_k_corrected = S_k_initial * (1 + α * O_k), where α is the speeding impact coefficient (e.g., set to 0.5). For events involving road condition defects, the avoidability percentage P_avoidable may be used to adjust the responsibility of relevant parties.

[0052] Normalization yields the final liability ratio: The corrected liability scores of all relevant parties are normalized to obtain the final liability ratio. That is, for subject k, its final liability ratio R_k = S_k_corrected / Σ_{all subjects j} S_j_corrected.

[0053] Specific numerical calculation examples: Suppose an accident involves driver A (vehicle) and road owner B. Analysis triggers two rules: Rule R1 (Driver speeding): Base weight W_R1_base = 0.7, simulation results show that driver A's speeding percentage O_A = 20%.

[0054] Rule R2 (Potential road surface defects due to poor maintenance by the property owner): Base weight W_R2_base = 0.6. Simulation analysis shows that even if the vehicle is not speeding, there is still a 30% probability of an accident occurring (i.e., avoidability P_avoidable = 70%).

[0055] Calculation process: Initial scores: S_A_initial = 0.7, S_B_initial = 0.6.

[0056] Application correction: Assuming the overspeed impact coefficient α = 0.5, then S_A_corrected = 0.7 * (1 + 0.5 * 0.2) = 0.7 * 1.1 = 0.77. For B, the avoidability P_avoidable = 70% indicates a significant contribution from the defect, which may be directly adopted or subject to similar correction. Here, it is simplified to S_B_corrected = 0.6 * (1.0 + 0.3) = 0.78 (assuming (1-P_avoidable) is used as part of the impact factor).

[0057] Normalized: Total score = 0.77 + 0.78 = 1.55. Final responsibility ratio: R_A = 0.77 / 1.55 ≈ 49.7%, R_B = 0.78 / 1.55 ≈ 50.3%.

[0058] Report generation steps: Integrate the results of responsibility allocation, the list of triggered rules, the index of key evidence, simulation parameters, and a detailed summary of the quantitative calculation process to form a structured and auditable judgment report.

[0059] III. Lightweight Road Right-of-Way Digital Twin and Visualization Module (300) This module serves as an "immersive window" for system-user interaction. It employs mainstream Web 3D technologies such as WebGL and loads an optimized, real-world 3D model generated by UAV oblique photogrammetry. The module (300) receives a "result data packet" from the decision engine (200).

[0060] In the 3D visualization interface, users can: (1) Time and space rewind: The entire process of the event is dynamically played through the time axis control bar. Vehicles and pedestrians move along the real trajectory, and the traffic lights change synchronously.

[0061] (2) Focus on responsibility: The system automatically highlights the main responsible party in a bright color (such as red) and displays their fault labels in real time.

[0062] (3) In-depth information mining: Click on any object in the scene, such as a car, a pothole or a section of guardrail, and the information panel on the right will immediately display its detailed information, including attributes, ownership, and role in this event.

[0063] (4) Simulation: Users can modify certain parameters in “simulation mode” (e.g., “if the pothole has been repaired” or “if the vehicle speed is reduced by 20%). The system can quickly recalculate and visualize the simulation results to help understand the impact of different factors on liability.

[0064] The resulting interactive 3D report can be shared with traffic police, insurance adjusters, parties involved, or municipal management departments via a secure URL link, making the professional liability determination results extremely intuitive and easy to understand, greatly improving communication efficiency and the credibility of the conclusions.

[0065] Workflow example: In a traffic accident, a car lost control and crashed into a guardrail while trying to avoid a pothole. After the system was activated: 1) A drone collected video and laser point cloud data of the scene and generated a chain of evidence; 2) The judgment engine identified the pothole (attributes: area 0.15㎡, depth 8cm), tracked the vehicle for speeding (speeding exceeded by 20% based on trajectory calculations), and found that the road section belonged to Company A and its maintenance was overdue; 3) Simulation analysis showed that within the standard reaction time, if there were no non-road defects, the vehicle had a 70% probability of avoiding the accident (i.e., avoidability P_avoidable=70%); 4) The rule base triggered "speeding" (rule R1). 5) The liability reasoning unit calculates based on the above quantitative model: the driver's initial score is 0.7, which is corrected to 0.77 after speeding; the property owner's initial score is 0.6, which is corrected to 0.78 after avoidability; after normalization, the driver is determined to bear approximately 49.7% of the responsibility, and the property owner, Company A, is determined to bear approximately 50.3% of the responsibility; 6) All processes, data and conclusions are clearly displayed in the three-dimensional digital twin scene in the form of highlights, animations, information panels and calculation process pop-ups.

[0066] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A road responsibility automatic determination system based on deep learning, characterized in that, include: The multi-source data collaborative acquisition and evidence collection module is used to synchronously collect multimodal data from the event scene and perform spatiotemporal alignment, and generate a tamper-proof evidence chain based on chain hashing and trusted timestamp technology; A decision engine integrating deep learning and rule-based reasoning, connected to the evidence collection module, is used to receive and parse the evidence chain, and includes: A multimodal perception and understanding unit is used to identify traffic participants, road facilities and environmental elements from the multimodal data, and output their status and spatiotemporal trajectory; The spatiotemporal scene reconstruction and reasoning unit is used to construct a unified spatiotemporal scene model that integrates dynamic trajectories, static road network information and ownership rules, and to perform event inference and quantitative analysis based on the model. A configurable rule knowledge base stores computer-executable logical rules that transform liability determination regulations into their equivalents. The responsibility reasoning and judgment unit is used to extract events from the spatiotemporal scene model, call the rule knowledge base for matching and reasoning, and output responsibility judgment conclusions and basis. The lightweight right-of-way digital twin and visualization module, connected to the determination engine, is used to integrate the responsibility determination conclusion and process data with the three-dimensional real-scene model to generate an interactive visualization report.

2. The system according to claim 1, characterized in that, The multi-source data collaborative acquisition and evidence collection module includes an integrated acquisition terminal deployed on a vehicle platform, drone, or roadside facility. The terminal has a built-in unified timing unit for aligning video frames, point cloud data, inertial measurement unit (IMU) data, and global navigation satellite system (GNSS) data within time windows. The module generates content hashes for each window of data in chronological order and concatenates the current content hash with the previous hash chain value to calculate a new chain value, forming a chain structure. The record containing the chain value is digitally signed using an end-security key and bound with a timestamp from a trusted third party before being written to a write-once-read-many (WORM) memory.

3. The system according to claim 1, characterized in that, The multimodal perception and understanding unit includes a multi-task neural network based on a shared feature extraction backbone. The neural network performs the following tasks in parallel: detecting, segmenting, and tracking traffic participants across frames; performing semantic segmentation and state recognition on lane lines, traffic signs, and traffic lights; and detecting and assessing the severity of road surface defects.

4. The system according to claim 1, characterized in that, The spatiotemporal scene reconstruction and reasoning unit performs the following operations: based on sensor calibration parameters, it maps the target trajectory and facility location output by the multimodal perception and understanding unit to a unified world coordinate system; it imports a high-precision digital map containing lane topology, right-of-way boundaries, and facility ownership information; it constructs a spatiotemporal scene model using a graph structure, where nodes represent physical or logical entities, and edges represent spatiotemporal, ownership, or rule relationships between entities; the unit also integrates a traffic flow simulation model, which is used to simulate the event occurrence process based on the spatiotemporal scene model and output collision dynamics parameters and avoidability analysis indicators.

5. The system according to claim 1, characterized in that, The rules in the configurable rule knowledge base are represented in the form of "condition-conclusion". The conditions are associated with the node attributes, edge relationships or event types in the spatiotemporal scene model, and the conclusions include the responsible entity identifier, the fault type and the weight coefficient. The knowledge base provides a graphical management interface that supports the editing, retrieval and priority configuration of rules.

6. The system according to claim 1, characterized in that, The workflow of the responsibility reasoning and judgment unit includes: Event extraction steps: Based on preset spatiotemporal and state thresholds, automatically identify potential violations or facility defects related events from the spatiotemporal scene model; Rule matching and triggering steps: Match the extracted event features with the conditions in the rule knowledge base to activate all applicable rules; Responsibility quantification step: Calculate the responsibility contribution of each relevant party by combining the weight coefficients of the triggered rules and the quantitative analysis indicators output by the spatiotemporal scene reconstruction and reasoning unit; Report generation steps: Integrate the results of responsibility allocation, the list of triggered rules, the index of key evidence, and a summary of the analysis process to form a structured judgment report.

7. The system according to claim 1, characterized in that, The lightweight right-of-way digital twin and visualization module is implemented through a browser-accessible Web 3D engine; the 3D reality model it loads is constructed from oblique photogrammetry point cloud data and optimized by level of detail (LOD); the module supports highlighting responsible parties in the 3D scene, dynamically replaying event processes, interactively querying the attributes and ownership information of any facility, and allowing users to adjust parameters to perform responsibility simulation.

8. A method for automatically determining road responsibility based on the system described in any one of claims 1-7, characterized in that, Includes the following steps: S1: Simultaneously acquire on-site data through the multi-source data collaborative acquisition and evidence collection module, and generate a tamper-proof evidence chain; S2: The evidence chain is analyzed through the judgment engine that integrates deep learning and rule reasoning to complete environmental perception, scene reconstruction and event analysis; S3: In the reconstructed spatiotemporal scene model, automated logical reasoning and responsibility quantification are performed based on a configurable rule knowledge base; S4: The judgment results are fused and presented interactively through the lightweight right-of-way digital twin and visualization module.

Images