Explanatory Accident Liability Determination Methods and Systems

Abstract

This application relates to the field of intelligent transportation technology, and provides an interpretable accident liability determination method and system. The method includes: acquiring multi-source accident data, then performing rule-based liability determination based on the multi-source data to generate rule-based determination results; and performing model-based determination based on the multi-source data to generate model-based determination results, wherein the model-based determination results are obtained by performing temporal analysis and spatial analysis on two neural network models respectively; comparing the rule-based determination results with the model-based determination results and outputting a difference result; when the difference result is greater than or equal to a preset difference threshold, performing interpretive analysis on the model-based determination results to output analysis results; and determining the determination result based on the analysis results. This method uses a hybrid decision-making approach combining rule-based determination and model-based determination, and through interpretive analysis, it can output the contribution of features in the multi-source data to the model-based determination results, thereby solving the problem of low reliability of the determination results.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation technology, and in particular to an interpretable method and system for determining accident liability. Background Technology

[0002] This system, based on multi-source data fusion for determining liability in traffic accidents, is used in commercial vehicle insurance claims scenarios. It collects multimodal data, including ADAS (Automatic Emergency Braking System) warnings, video images, vehicle CAN bus data, and GPS trajectory data, through onboard sensors. Liability is then determined using rule-based reasoning or data-driven models. To handle a large volume of accidents, the system must balance efficiency, scenario adaptability, result reliability, and continuous evolution capabilities.

[0003] Some solutions utilize expert rule systems built upon traffic regulations to achieve rapid liability matching; others employ deep learning models to perform end-to-end liability classification on multimodal data; still others employ probability calculation systems built upon Bayesian networks and causal reasoning to output liability probabilities and causal paths.

[0004] Rule-based systems struggle to handle complex, undefined scenarios; machine learning models suffer from opaque decision-making processes and unreliable results; and probabilistic reasoning systems rely heavily on prior knowledge and lack self-optimization capabilities, resulting in low credibility of generated judgments. Summary of the Invention

[0005] This application provides an interpretable accident liability determination method and system to address the problem of low reliability of determination results.

[0006] Firstly, this application provides an interpretable method for determining liability in an accident, including:

[0007] Acquire multi-source data on the incident;

[0008] Based on the multi-source data, a rule responsibility determination is performed to generate a rule determination result; and based on the multi-source data, a model determination is performed to generate a model determination result, wherein the model determination result is obtained by performing temporal analysis and spatial analysis on two neural network models respectively.

[0009] Compare the rule-based judgment results with the model judgment results to output the difference results;

[0010] When the difference result is greater than or equal to a preset difference threshold, an interpretive analysis is performed on the model judgment result to output the analysis result, which is the contribution of the features in the multi-source data to the model judgment result;

[0011] Based on the analysis results, a judgment result is determined, which is either a rule-based judgment result or a model-based judgment result.

[0012] In some feasible embodiments, the step of performing rule responsibility determination based on the multi-source data to generate rule determination results includes:

[0013] The multi-source data is processed using a rule engine;

[0014] Based on a pre-set traffic rule knowledge base, logical judgments are performed to generate judgment results;

[0015] Based on the judgment result, the rule judgment result is generated.

[0016] In some feasible embodiments, the step of performing model determination based on the multi-source data to generate model determination results includes:

[0017] Acquire a temporal neural network and a convolutional neural network; the temporal neural network is a long short-term memory network.

[0018] The temporal neural network is used to process the temporal data in the multi-source data to extract temporal features;

[0019] The image data from the multi-source data is processed by the convolutional neural network to extract spatial features;

[0020] The temporal features and spatial features are fused to generate fused features;

[0021] The model determination result is generated based on the fusion features.

[0022] In some feasible embodiments, the timing data includes one or more combinations of vehicle speed data, turn signal data, brake pedal depth data, steering wheel angle data, and positioning trajectory data;

[0023] The step of processing the time-series data in the multi-source data through the time-series neural network to extract time-series features includes:

[0024] Preprocessing is performed on the vehicle speed data, turn signal data, brake pedal depth data, steering wheel angle data, and positioning trajectory data to generate preprocessed timing data;

[0025] The preprocessed time series data is input into the Long Short-Term Memory network to extract hidden states that characterize vehicle motion patterns and driving behavior features through the Long Short-Term Memory network.

[0026] The hidden state is determined as the temporal feature.

[0027] In some feasible embodiments, the image data includes video data from multiple vehicle-mounted cameras;

[0028] The step of processing image data from the multi-source data using the convolutional neural network to extract spatial features includes:

[0029] Extract keyframe images from the video data;

[0030] Preprocessing is performed on the keyframe images to generate preprocessed images;

[0031] The preprocessed image is input into the convolutional neural network to extract image features from the bottom layer to the top layer through the multi-layer convolution and pooling operations of the convolutional neural network;

[0032] The top-level image features are mapped into semantic information representing the spatial relationships of traffic scenes;

[0033] The semantic information is parsed to generate spatial features, which represent vehicle positional relationships, lane line structure, and road environment information.

[0034] In some feasible embodiments, performing interpretive analysis on the model determination results to output analysis results includes:

[0035] Extract features from the multi-source data;

[0036] The contribution of the feature to the model's judgment result is calculated using an interpretive analysis method based on game theory.

[0037] The analysis results are generated based on the stated contribution.

[0038] In some feasible embodiments, determining the judgment result based on the analysis results includes:

[0039] Determine whether the analysis results conform to preset traffic rules and common sense judgment standards;

[0040] If the conditions are met, the model determination result is determined as the determination result;

[0041] If the conditions are not met, the result of the rule determination shall be determined as the determination result.

[0042] In some feasible embodiments, the method further includes:

[0043] Cases where the difference result is greater than or equal to the preset difference threshold are marked as conflict cases;

[0044] The multi-source data of the aforementioned conflict cases are added to the training dataset to obtain an updated dataset;

[0045] The updated dataset is used to train both a temporal neural network and a convolutional neural network.

[0046] In some feasible embodiments, the multi-source data includes at least one of the following: driving assistance event data, braking system event data, in-vehicle video data, bus data, and positioning data;

[0047] Get the first parameter, the second parameter, the third parameter, the fourth parameter, and the fifth parameter;

[0048] The first parameter is to extract the warning event type and trigger time from the driving assistance event data; the second parameter is to obtain the braking activation state and braking intensity from the braking system event data; the third parameter is to identify the behavior of traffic participants and the status of road signs from the vehicle video data; the fourth parameter is to parse the vehicle operating status parameters from the bus data; and the fifth parameter is to calculate the vehicle movement trajectory and relative position relationship from the positioning data.

[0049] The first parameter, the second parameter, the third parameter, the fourth parameter, and the fifth parameter are matched with preset traffic rule conditions to generate a matching result;

[0050] Based on the matching results, a rule determination result is generated.

[0051] Secondly, this application provides an interpretable accident liability determination system, comprising:

[0052] The acquisition unit is configured to acquire multi-source data about the incident.

[0053] The result determination unit is configured to perform rule responsibility determination based on the multi-source data to generate rule determination results; and to perform model determination based on the multi-source data to generate model determination results.

[0054] The analysis unit is configured to compare the rule-based determination result with the model-based determination result to output a difference result, wherein the model-based determination result is obtained by performing temporal analysis and spatial analysis on two neural network models respectively; and when the difference result is greater than or equal to a preset difference threshold, to perform interpretive analysis on the model-based determination result to output an analysis result, wherein the analysis result is the contribution of features in the multi-source data to the model-based determination result; and based on the analysis result, to determine a determination result, wherein the determination result is a rule-based determination result or a model-based determination result.

[0055] As can be seen from the above technical solutions, this application provides an interpretable accident liability determination method and system. The method includes: acquiring multi-source accident data, then performing rule-based liability determination based on the multi-source data to generate a rule-based determination result; and performing model-based determination based on the multi-source data to generate a model-based determination result, wherein the model-based determination result is obtained by performing temporal analysis and spatial analysis on two neural network models respectively; comparing the rule-based determination result with the model-based determination result to output a difference result; when the difference result is greater than or equal to a preset difference threshold, performing interpretive analysis on the model-based determination result to output an analysis result; and determining a determination result based on the analysis result, wherein the determination result is either a rule-based determination result or a model-based determination result. This method uses a hybrid decision-making approach combining rule-based determination and model-based determination, and through interpretive analysis, it can output the contribution of features in the multi-source data to the model-based determination result, thereby solving the problem of low reliability of the determination result. Attached Figure Description

[0056] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart illustrating the interpretability accident liability determination method provided in the embodiments of this application;

[0058] Figure 2 This is a schematic diagram of multi-source output transmission between in-vehicle equipment and vehicle networking platform provided in an embodiment of this application;

[0059] Figure 3 A schematic diagram of the rule determination process provided in the embodiments of this application;

[0060] Figure 4 This is a schematic diagram of the model determination process provided in an embodiment of this application;

[0061] Figure 5 This is a schematic diagram of the model update process provided in an embodiment of this application. Detailed Implementation

[0062] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following examples do not represent all embodiments consistent with this application.

[0063] In determining liability for accidents, while preset traffic rules can be used, they are insufficient for handling complex accident liability determinations. Machine learning and multimodal large models can also be employed, but these present challenges in terms of inference time, accuracy, and cost. Furthermore, current technologies can only output "what"—the proportion of liability—without explaining "why." This leads to a lack of understanding and trust among relevant parties, making the results difficult to accept in judicial and insurance claims scenarios.

[0064] Whether it's a human-based experience system or a pre-set rule system, its knowledge base and capability boundaries are relatively static. It cannot automatically learn and evolve from the cases it has processed, nor can it cope with future traffic scenarios and accident patterns. As time goes by, its effectiveness will gradually decline.

[0065] To address the aforementioned issues, some embodiments of this application provide an interpretable accident liability determination method. This method integrates pre-set traffic rules and machine learning-based hybrid decision-making, which can compensate for the shortcomings of pre-set traffic rules in terms of scenarios and the excessive time or cost of model recognition and reasoning. It leverages the efficiency of the rule engine and the generalization ability of the machine learning model, enabling it to quickly process a large number of routine accident cases while also handling complex situations, achieving a balance between efficiency and accuracy. Furthermore, by introducing interpretability technologies such as SHAP (SHapley Additive exPlanations), the results are linked to traffic rules, providing clear and common-sense reasoning for each determination, making the automated accident determination process transparent, thereby improving credibility.

[0066] Among them, SHAP is a method from game theory to explain the prediction results of any machine learning model. It decomposes complex model predictions and assigns them to each input feature, telling the model "why" it makes such a decision in a quantitative and comparable form.

[0067] like Figure 1 As shown, the method includes the following steps:

[0068] S100: Acquire multi-source data on the accident.

[0069] Multi-source data is a collection of various types of data collected from vehicles and their onboard systems during an accident, which may originate from different sensors and control units. In some embodiments, multi-source data includes at least one of the following: driver assistance event data, braking system event data, onboard video data, bus data, and location data. These data are correlated in time and space, and together constitute a chain of evidence for reconstructing the accident process and determining liability.

[0070] Among them, driver assistance event data is event records generated by ADAS, braking system event data is intervention signals triggered by the Automatic Emergency Braking System (AEBS), vehicle video data is video streams continuously recorded by vehicle cameras, bus data is real-time vehicle operating parameters transmitted through the Controller Area Network (CAN) bus, and positioning data is trajectory coordinates provided by the Global Positioning System (GPS) receiver.

[0071] When a traffic accident occurs, various data acquisition devices deployed on the vehicle will be triggered, uploading driving assistance event data, braking system event data, in-vehicle video data, bus data, and positioning data for a period of time before and after the event to the data processing platform.

[0072] For vehicles equipped with ADAS or AEBS systems, which possess the ability to warn of abnormal situations and perform emergency braking, it is necessary to upload data before and after an event to the vehicle-to-everything (V2X) platform in real time to construct an accident evidence chain. ADAS or AEBS systems can provide driver assistance warnings such as Forward Collision Warning (FCW), Lane Departure Warning (LDW), Headway Monitoring and Warning (HMW), Blind Spot Detection (BSD), Autonomous Emergency Braking (AEB), and Driver Monitoring System (DMS).

[0073] In some embodiments, a first parameter, a second parameter, a third parameter, a fourth parameter, and a fifth parameter are obtained; wherein, the first parameter is the type of warning event and the trigger time extracted from the driving assistance event data, the second parameter is the braking activation state and braking intensity obtained from the braking system event data, the third parameter is the behavior of traffic participants and the status of road signs identified from the vehicle video data, the fourth parameter is the vehicle operating status parameters parsed from the bus data, and the fifth parameter is the vehicle trajectory and relative position relationship calculated from the positioning data;

[0074] The first, second, third, fourth, and fifth parameters are matched with preset traffic rule conditions to generate a matching result; based on the matching result, a rule determination result is generated.

[0075] The CAN bus can collect vehicle information, including vehicle speed, left turn signal, right turn signal, hazard light signal, brake light signal, gear position signal, accelerator pedal depth, brake pedal depth, vehicle mileage, steering wheel angle, etc., which is the fourth parameter. The GPS module can collect vehicle position, heading angle, speed, elevation, positioning status, and reporting time, which is the fifth parameter.

[0076] When a warning event occurs, detailed collision information, including obstacle type, estimated collision time, longitudinal relative distance, and following distance, is acquired and uploaded to the vehicle network platform in real time. At the same time, multiple camera video feeds from the ADAS or AEBS system at the moment of the warning event are captured by the vehicle's video system and uploaded to the vehicle network platform.

[0077] like Figure 2 As shown, for example, the ADAS or AEBS system collects event data and timestamps before and after the warning event, the video camera system collects video data and timestamps before and after the warning event, four cameras (front, rear, left, and right) record simultaneously, the CAN bus acquires real-time data such as vehicle speed, steering angle, turn signals, and brake pedal travel, the GPS module records the vehicle's real-time position and trajectory, and the on-board data acquisition module sends the collected data to the vehicle networking platform in accordance with the JTT808 protocol format.

[0078] After parameter extraction, rule matching is performed. The specific values ​​of the first to fifth parameters are logically compared one by one with the conditional part of each rule in the traffic rule knowledge base. For example, a rule regarding rear-end collision liability requires matching the following facts: "the rear vehicle's speed is greater than the front vehicle's speed in the fourth parameter," "the distance between the two vehicles is less than the safe distance in the fifth parameter," and "the front vehicle's braking intensity is greater than the threshold in the second parameter." The matching process outputs a matching result.

[0079] If any rule's condition is met, the responsibility determination specified in the conclusion section is output. If multiple rules are triggered, the final conclusion is determined based on rule priority or conflict resolution strategy. If no rules are fully met, an "undetermined" flag can be output, or another reasoning process can be triggered.

[0080] To improve accuracy and completeness, multi-source data needs to undergo preprocessing steps after acquisition. These steps include aligning data from different sources on a timeline, removing outliers with obvious errors, and standardizing the data format.

[0081] For example, a unified calibration is performed based on the timestamp, then abnormal data is removed, missing values ​​are filled in, key frames are identified from the video, target recognition is performed using the YOLO algorithm, and key parameters are extracted from the CAN data.

[0082] In some embodiments, multi-source data is mapped to a unified spatiotemporal coordinate system (x, y, t) for location and time as follows. ＇ y ＇ , t ＇ ):

[0083] ;

[0084] Where θ is the vehicle heading angle, and (x0, y0, t0) is the accident reference point.

[0085] Based on a unified spatiotemporal coordinate system, data from various data sources are fused to generate fused target trajectories and events, constructing a spatiotemporal event chain to reproduce the accident scenario.

[0086] S200: Based on multi-source data, perform rule responsibility determination to generate rule determination results; and, based on the multi-source data, perform model determination to generate model determination results, wherein the model determination results are obtained by performing temporal analysis and spatial analysis on two neural network models respectively.

[0087] For the processed multi-source data, rule-based liability determination and model-based determination are performed separately. Rule-based liability determination is a process of logical reasoning and judgment of accident liability based on preset, digitized traffic regulations. Model-based determination is a process of using machine learning models to perform in-depth analysis of multi-source data to infer accident liability.

[0088] like Figure 3 As shown, in some embodiments, a rule engine is used to process the multi-source data; then, based on a preset traffic rule knowledge base, logical judgments are performed to generate judgment results; and based on the judgment results, the rule determination results are generated.

[0089] The rule-based liability determination is performed using the Drools rule engine, which can handle "if-then" conditional rules. The determination is based on a traffic rule knowledge base, which decomposes road traffic safety regulations and transforms them into a collection of computer-executable logical rules. Each rule consists of a condition part and a conclusion part. The condition part defines the combination of data states required to trigger the rule, and the conclusion part represents the liability determination to be made when the conditions are met, enabling rapid preliminary liability assessment for common accident types.

[0090] The traffic rules knowledge base in Drools is specifically represented by one or more .drl rule files. These files are containers for storing all digitized traffic rules. Each rule is written using the Drools rule language, clearly defining the conditions that must be met to trigger the rule and the actions that should be taken once the conditions are met.

[0091] Logical judgment involves the rule engine matching the input multi-source data with the rule conditions in the traffic rule knowledge base and deriving a conclusion based on the matching results. The execution of logical judgment follows pattern matching; the inference engine traverses all rules in the traffic rule knowledge base, checking whether the condition part of each rule is satisfied by the data in the current working memory. For a complex liability determination rule, its condition part may consist of multiple sub-conditions connected by logical operators. The logical judgment process needs to evaluate these sub-conditions one by one and calculate the final condition truth value according to the operators.

[0092] The output of the rule-based liability determination is the rule determination result, which is a preliminary conclusion about the proportion of responsibility of each party in an accident. It can be a clear division of responsibility, such as full responsibility, primary responsibility, equal responsibility, secondary responsibility, or no responsibility, or it can be a quantified percentage of responsibility.

[0093] The rule engine generates a rule determination result based on the judgment result. If the logical judgment produces only one judgment result, that result can be output as the rule determination result. When there are multiple judgment results and their conclusions conflict with each other, the rule engine can filter or integrate the final conclusion, i.e., the rule determination result, from multiple potential judgment results based on the priority of the rules, the weight of the rules, or the strategy of the latest conclusion overriding the old conclusion.

[0094] Model-based judgment is derived from the latent patterns and complex relationships learned from historical accident data. The model-based judgment result is obtained by performing temporal and spatial analyses on two neural network models respectively. In other words, it is not generated by a single model, but by a neural network that processes temporal data and a neural network that processes spatial image data working together and fusing their analysis results. Similar to the rule-based judgment result, the model-based judgment result is also a conclusion about the allocation of responsibility.

[0095] Among them, the neural network for processing time-series data is a time-series neural network, and the neural network for processing spatial image data is a convolutional neural network (CNN). The time-series neural network is a long short-term memory (LSTM) network. The model receives time-series data and spatial image data simultaneously through a dual-channel data processing method.

[0096] like Figure 4 As shown, in some embodiments, a temporal neural network and a convolutional neural network are obtained; the temporal data in the multi-source data is processed by the temporal neural network to extract temporal features; the image data in the multi-source data is processed by the convolutional neural network to extract spatial features; the temporal features and the spatial features are fused to generate fused features; and the model determination result is generated based on the fused features.

[0097] LSTM networks analyze historical vehicle behavior sequences, capturing motion patterns and dynamic changes to obtain temporal features. Temporal features are high-level abstract representations extracted by temporal neural networks from multi-source input data, characterizing the behavioral patterns and states of an object over time. In vehicle accident analysis, temporal features can encode dynamic behavioral patterns such as acceleration trends, braking response delays, steering smoothness, and following distance change rates.

[0098] CNN networks process bird's-eye view scene images, extracting road structure and spatial relationship features to obtain spatial features. Spatial features are high-level abstract representations extracted by convolutional neural networks from multi-source input image data, characterizing the structure, shape, position, and geometric relationships of objects in the scene. In vehicle accident analysis, spatial features can encode the precise position of vehicles at the moment of the accident (whether they crossed the lines), the relative distance and angle between vehicles, the clarity and type of lane lines, the presence of traffic signs, and the road topology.

[0099] After the two features are fused, they undergo nonlinear transformation and decision reasoning through a fully connected layer to output a quantitative result of the three-party responsibility ratio.

[0100] It is understandable that feature fusion is not a simple feature concatenation, but rather considers the interactions and contribution weights between features to ensure that the fused features retain the most useful information from each modality while suppressing redundant fusion methods. Different weight coefficients and confidence calculation models can be set for different multi-source data. For example, for ADAS and AEB events, the weight coefficient can be 0.8, and the confidence calculation model can be a Bayesian network.

[0101] By combining temporal dynamic analysis with spatial static assessment, both the process characteristics of accident development and the scene factors at the moment of the incident can be taken into account.

[0102] Specifically, a traffic accident is an eruption of a dynamic process within a static scene. These two types of information are fundamentally different and require different neural network architectures to efficiently extract their essential features. The nature of the data modality determines the model architecture. Traffic accident liability itself is a spatiotemporally coupled event, which can unify the behavioral process with a snapshot of the scene.

[0103] LSTM is a recurrent neural network designed for sequential data. It has internal gating mechanisms, such as forget gate, input gate, and output gate, which can capture long-term dependencies. Using LSTM to analyze time series data can more directly analyze the process and behavior of accidents.

[0104] In some embodiments, timing data includes one or more combinations of CAN bus and GPS data; CAN bus data includes, for example, vehicle speed data, turn signal data, brake pedal depth data, and steering wheel angle data; GPS data includes, for example, positioning trajectory data. Data at each point in time, such as vehicle speed and steering wheel angle, are related to previous and subsequent states, exhibiting continuity and dependency. LSTM is used to answer questions such as "How did it develop to this point?", and "Three seconds before the accident, were the brake lights of the vehicle in front already on? How long were they on? Did the vehicle exhibit any sudden or abnormal steering behavior?" etc.

[0105] The vehicle speed data, turn signal data, brake pedal depth data, steering wheel angle data, and positioning trajectory data are preprocessed to generate preprocessed time-series data; the preprocessed time-series data is input into the long short-term memory network to extract hidden states that characterize vehicle motion patterns and driving behavior features; and the hidden states are determined as the time-series features.

[0106] Understandably, different time-series data undergo different preprocessing procedures during the preprocessing process. For continuous numerical data such as vehicle speed data and brake pedal depth data, standardization or normalization is performed to scale them to zero mean and unit variance, or to map them to the zero-to-one interval, in order to eliminate the influence of dimensions and accelerate network convergence. For discrete state data such as turn signal data, they are converted into one-hot encoding or integer label encoding so that the neural network can process them effectively. For all data, outlier detection and smoothing filtering are performed again to further improve data quality and ensure that the signal input to the network is clean and reliable. The processed data is organized into a regular multidimensional tensor by time step, i.e., preprocessed time-series data.

[0107] Preprocessed temporal data is input into a Long Short-Term Memory (LSTM) network for forward propagation computation. The network processes the input sequence sequentially time-step by time. At each time step, the three gating mechanisms of the LTM network work together. The forget gate calculates which old information needs to be discarded from the cell state based on the current input and the previous hidden state. The input gate decides which new candidate information (generated by the current input and the previous hidden state through a tanh activation function) to add to the cell state. The cell state is updated based on the outputs of the forget gate and the input gate. Finally, the output gate calculates and outputs the hidden state for the current time step based on the updated cell state, the current input, and the previous hidden state. The hidden state includes the sequence information up to the current time step. The network processes the entire sequence sequentially, and the hidden state output at the final time step represents the crucial information of the entire incident.

[0108] The hidden state vector generated at the final time step of the Long Short-Term Memory network is then used as the temporal feature to be output in this feature extraction.

[0109] In some embodiments, the image data includes video data from multiple vehicle-mounted cameras; keyframe images are extracted from the video data; preprocessing is performed on the keyframe images to generate preprocessed images; the preprocessed images are input into the convolutional neural network to extract image features from the bottom layer to the top layer through multi-layer convolution and pooling operations of the convolutional neural network; the top layer image features are mapped to semantic information representing spatial relationships in a traffic scene; the semantic information is parsed to generate spatial features, which represent vehicle position relationships, lane line structures, and road environment information.

[0110] Keyframe images are static images extracted from continuous video data that characterize a specific stage of an accident or possess specific features. Keyframe images are acquired through algorithms, which may include methods based on scene change detection to automatically extract the frame when video content changes, such as a vehicle suddenly cutting in or a collision occurring. Algorithms may also include methods based on fixed time interval sampling to extract images at specific frequencies before and after the accident timestamp. The acquired keyframe images contain crucial instantaneous data regarding the relative positions of vehicles involved in the accident, the visibility of lane markings, traffic signal status, and the road environment.

[0111] For example, based on the accident trigger signal, such as the collision warning time issued by ADAS, a time window for analysis is determined. Within the time window, video frames are traversed using content-based analysis methods to calculate the difference or saliency between consecutive frames. When the difference exceeds a threshold, the frame is determined to carry key change information. Alternatively, images can be sampled at a specific frequency at fixed time intervals, such as several seconds before and after the accident, to supplement the extraction of a set of keyframe images.

[0112] To eliminate the differences between different preprocessed images, various preprocessing methods are used, such as standardization and enhancement operations. Specifically, all input images are adjusted to a fixed size required by the convolutional neural network, pixel values ​​are normalized to normalize pixel intensity values ​​from an integer range to a specific interval, color channels are adjusted to perform color consistency correction, and data augmentation can also be included to increase data diversity by rotating, translating, or adjusting contrast.

[0113] The preprocessed image is input into a convolutional neural network for forward propagation computation to extract multi-level feature representations. The first convolutional kernel slides across the input image, generating preliminary feature maps by calculating weighted sums of local regions and applying nonlinear activation functions. These feature maps respond to basic edge, color, and texture patterns. Other convolutional layers further combine these low-level feature maps to form more complex pattern representations, such as wheel shapes, headlight combinations, and vehicle outlines. Pooling layers are interspersed between convolutional layers, reducing spatial dimensionality and providing some translation invariance by downsampling the feature maps. The deeper the network, the larger its receptive field, and the more abstract and high-level the extracted features. Through convolution and pooling operations, the network progressively abstracts image features representing patterns of varying complexity from the bottom to the top layers.

[0114] The most abstract image features output from the top layer of a convolutional neural network are mapped to semantic information representing spatial relationships in a traffic scene. While these high-level features are rich in information, they are still distributed representations and lack explicit semantics. The mapping process is implemented through specific structures or post-processing modules at the network's ends. For example, feature maps can be input into a region proposal network, which then uses classification and regression to obtain the category and location of specific objects in the image. Alternatively, feature maps can be input into a fully connected layer after global average pooling, where each output node corresponds to a predefined semantic concept. In this way, abstract feature vectors are transformed into understandable semantic descriptions, such as "Vehicle A is located at the northeast corner of the intersection," "Vehicle B is crossing the solid line," and "The traffic light is red."

[0115] The semantic information is then parsed to generate spatial features. This parsing process involves the quantification and structured encoding of the semantic information. For example, from the semantic description "vehicle A is located in the center of the lane," the pixel coordinates of the center point of vehicle A are parsed and converted into real-world coordinates relative to the vehicle's coordinate system. From "vehicle B is crossing the solid line," parameters such as the distance between vehicle B and the solid line of the lane and the angle of crossing are parsed. These parameters are organized into structured numerical vectors or tensors, i.e., spatial features.

[0116] CNNs, through convolutional kernels and pooling operations, excel at extracting local features from images, such as edges and corners, and progressively combining them into more complex patterns, such as wheels, vehicles, and lanes, exhibiting invariance to translation and rotation. Using CNNs to analyze spatial data, the scene and relationships of an accident can be clearly analyzed. Key information can be captured in the spatial local relationships and hierarchical structures between pixels, such as lane lines, vehicle positions, and road edges. CNNs can answer questions like "What was the scene like at the moment of the incident?", such as "Did the vehicle cross the line? In which lane? What was the road environment like? What were the positions of the participants?"

[0117] The combination of LSTM and CNN is also based on the advantages of separating, extracting and fusing the two features. Combining the two can decouple the features, making the information purer and avoiding the feature overload and interference that occurs when raw data from different modalities are directly mixed and input into a model. LSTM learns behavioral patterns, and CNN learns spatial relationships, and finally the feature representations of the two domains are obtained.

[0118] LSTM can better capture temporal patterns such as "premature deceleration," while CNN can more accurately identify spatial states such as "vehicles crossing lane lines." It also offers greater interpretability; the architecture itself provides initial interpretability by analyzing which time point the LSTM focuses on, and by visualizing the regions the CNN focuses on in the image, providing a foundation for subsequent analysis.

[0119] Under the probability determination of accident responsibility, the combination of LSTM and CNN can achieve the dialectical unity of the accident process and the accident scene. This is because the determination of accident responsibility is not a static snapshot problem, but a dynamic process problem that requires answering the two questions "Where was he at the time?" and "What did he do before?"

[0120] Using LSTM and CNN, we can quantify the "right of way" and "duty to drive safely" when determining liability. The output of CNN can clearly show the position of each party at the moment of the accident (whether they are in their own lane), which is directly related to the division of right of way. The output of LSTM reveals the behavior of each party before the accident (whether they were speeding, changing lanes illegally, or avoiding danger in time), which is directly related to whether they have fulfilled their duty to drive safely. After fusion, we can make complex judgments that comply with traffic regulations, such as "although the car in front suddenly changed lanes (space violation detected by CNN), the car behind followed too closely and did not brake in time (behavioral negligence detected by LSTM). Therefore, we can determine that both parties bear equal responsibility."

[0121] It can also enhance the ability to judge complex scenarios. In many accidents, it is not possible to accurately determine responsibility based on just one frame of footage or just behavioral data. The fused model can combine the facts of the collision in space with the behavioral faults in time sequence to give a fairer proportion of responsibility. For example, it can determine that the driver of the vehicle that cut in is mainly responsible. This is an effect that a single modal model cannot achieve.

[0122] Furthermore, it outputs a more refined probability distribution. The final output of the responsibility distribution (of this party, the other party, and the third party) is derived from the weighted decision of process factors and on-site factors. This makes the probability distribution no longer a vague guess, but supported by multi-dimensional data. For example, it can clearly show that of the 70% of the responsibility of this party, 50% is due to following too closely (temporal behavior) and 20% is due to the failure to avoid to the right in time (spatial opportunity).

[0123] In this embodiment, when different conclusions are reached based on rule-based responsibility determination and model determination, one of them will not be arbitrarily selected. Instead, a third-party interpreter will be introduced to audit and evaluate which result is more credible. The ultimate goal is to pursue the fairness and interpretability of the decision, rather than just the accuracy.

[0124] In determining liability for accidents, the principle of interpretability should be followed. By leveraging the logical transparency of rule-based liability determination and the clarity of decision-making basis of machine learning models, priority should be given to results whose decision-making process is more interpretable and whose feature contributions are more in line with traffic rules and common sense, thereby achieving a balance between technical complexity and the fairness and reasonableness of liability determination.

[0125] S300: Compare the rule-based decision results with the model-based decision results to output the difference results.

[0126] The difference result is a quantitative representation of the consistency between the rule-based decision and the model decision. It is calculated using a comparison function, which can calculate the Euclidean distance, absolute difference, or percentage difference between the two result vectors. The difference result is a numerical value used to measure whether there is a discrepancy between the rule-based logical judgment and the data-driven model prediction, and the magnitude of the discrepancy.

[0127] The comparison process involves calculating the degree of difference between two result vectors. If the difference is less than a preset difference threshold, the representation rule and the model's judgment are basically consistent, and either result or their weighted average can be used as the final output, at which point the process ends. However, when the difference is greater than or equal to the preset difference threshold, it indicates that the current accident case may be in a fuzzy area of ​​the rule, or that the model has learned a non-intuitive pattern, requiring further clarification.

[0128] S400: When the difference result is greater than or equal to the preset difference threshold, perform interpretive analysis on the model judgment result to output the analysis result, which is the contribution of the features in the multi-source data to the model judgment result.

[0129] The preset difference threshold is a pre-defined threshold value. When the difference result is greater than or equal to this threshold, the discrepancy between the representation rule and the model has reached a level that requires further review. In this embodiment, the preset difference threshold is 20%. When the difference result is greater than or equal to 20%, interpretive analysis is performed on the model's judgment result to output the analysis result.

[0130] Interpretive analysis, in particular, is the decision-making logic employed when the predictions of a machine learning model need to be understood and trusted. When there is a discrepancy between the model's decision and the rule-based decision, interpretive analysis is required to transform the model's decision-making process into a human-understandable basis for decision-making.

[0131] In this embodiment, the interpretive analysis, based on game theory principles, can fairly allocate the contribution of each input feature to the final prediction result.

[0132] Contribution is an analytical result output from interpretive analysis. It is a quantitative metric used to measure the magnitude and direction of the influence of each feature in multi-source data on the model's decision. Contribution can be expressed numerically; a positive contribution indicates that the feature tends to increase the proportion of responsibility for a particular party, while a negative contribution indicates that the feature tends to decrease the proportion of responsibility for a particular party.

[0133] In some embodiments, features are extracted from the multi-source data; the contribution of the features to the model's decision result is calculated using a game theory-based interpretive analysis method; and the analysis result is generated based on the contribution.

[0134] In accident liability determination scenarios, features can include numerical features, such as the mean and variance of vehicle speed, the depth of the brake pedal, and the time distance to the vehicle in front. Features can also include categorical features, such as the status of the turn signal, the position of the lane, and the color of the traffic light.

[0135] Interpretive analysis methods based on game theory aim to explain the prediction results of machine learning models. They originate from the principle of fair distribution of benefits in cooperative game theory, treating the model's prediction results as the benefits generated by the collaboration of all input features, and fairly distributing the contribution of each feature to this benefit through a mathematical framework.

[0136] In this embodiment, the game theory-based interpretive analysis method is SHAP, which allocates contribution by calculating the average of the marginal contributions of a feature appearing in all possible subsets of features.

[0137] like Figure 5 As shown, for cases where the difference result is greater than or equal to the preset difference threshold, in some embodiments, cases where the difference result is greater than or equal to the preset difference threshold are marked as conflict cases; the multi-source data of the conflict cases are added to the training dataset to obtain an updated dataset; and a temporal neural network and a convolutional neural network are trained using the updated dataset.

[0138] A conflict case means that the current accident situation is on the edge of the rule system, or that the model has learned a potential pattern that is not explicitly covered by the rules. When the difference is greater than or equal to a preset difference threshold, the multi-source data, the two judgment results, and the interpretability analysis report of the accident are marked as a conflict case.

[0139] Next, multi-source data of all labeled conflict cases are added to the original training dataset, initiating the model retraining process. The latest updated dataset is then obtained, and the temporal neural network and convolutional neural network are retrained using this data. The training process employs deep learning algorithms, aiming to minimize the error between the model's predictions and the true labels (the final judgment result after arbitration or expert annotation), iteratively updating the network's weight parameters. Through retraining, the model can learn from previous misjudgments or cases of disagreement, adjusting its internal decision boundaries to make more accurate predictions when encountering similar scenarios in the future.

[0140] S500: Based on the analysis results, determine the judgment result, which is either a rule-based judgment result or a model-based judgment result.

[0141] SHAP analysis uses the contribution of each feature data point in a machine learning model to the determination of liability. SHAP analysis can provide local interpretability of the model's prediction, revealing the basis for the model's decision. If the SHAP analysis shows that the model's decision-making basis is very clear and consistent with common traffic sense—for example, if the SHAP shows that the model's determination of full liability is mainly due to the significant contribution of the feature "the other vehicle ran a red light"—then the model's result is highly interpretable.

[0142] If SHAP shows that the model's decision depends on some difficult-to-understand or irrelevant features, such as the model's judgment depending on an irrelevant sensor noise, then even if the model is confident, the result is unreliable. In this case, the system will roll back to the result of the rule engine, which has clear logic and well-defined rules.

[0143] Similarly, for models with poor interpretability, the accident data will also be included in the model's training set to enable model self-evolution and model updates.

[0144] The judgment result represents the conclusion on the allocation of responsibility for the accident. It is a selective output based on the contribution of the interpretive analysis and a credibility assessment of the preliminary results. If the analysis shows that the model's decision is based on clear and common sense, the judgment result is the model's judgment result; otherwise, it is the rule-based judgment result.

[0145] In some embodiments, it is determined whether the analysis result conforms to preset traffic rules and common sense judgment standards; if it conforms, the model judgment result is determined as the judgment result; if it does not conform, the rule judgment result is determined as the judgment result.

[0146] The combination of pre-defined traffic rules and common-sense judgment standards in written traffic regulations and unwritten social consensus on driving is used to measure the rationality and acceptability of driving behavior or decision-making. This includes not only explicit provisions in road traffic safety laws, such as regulations on right-of-way, safe distance, and signal use, but also general perceptions formed based on long-term driving practice, such as defensive driving principles and reasonable predictions and reaction expectations in specific scenarios.

[0147] The analysis results, specifically the contribution values ​​generated by interpretive analysis, are compared with preset traffic rules and common-sense judgment standards. This comparison process is based on rule-based reasoning. By comparing features with high contribution values ​​(whether positive or negative), it is determined whether the driving behaviors or scenario states represented by these features clearly violate traffic regulations, or whether their influence weights conform to the general cognitive logic of human drivers. For example, it determines whether the high-contribution feature of "running a red light" truly corresponds to an absolutely prohibited behavior in regulations, or whether assigning a high contribution value to "sensor noise" constitutes an unreasonable decision that violates common sense.

[0148] If the analysis results conform to preset traffic rules and common sense judgment standards, the model's judgment result is determined as the final judgment result. In other words, although its initial conclusion differs from the rule engine, the model's decision-making rationale is deemed sufficient after review. If the analysis results do not conform to preset traffic rules and common sense judgment standards, the rule-based judgment result is determined as the final judgment result. This means that the model's decision-making basis is deemed unreliable or unreasonable, possibly due to model bias in a specific scenario or insufficient training. In this case, the model reverts to a logically clear rule-based judgment result derived from regulations to improve accuracy.

[0149] Based on the above-described method for determining liability in an interpretable accident, some embodiments of this application also provide a system for determining liability in an interpretable accident, including:

[0150] The acquisition unit is configured to acquire multi-source data about the incident.

[0151] The result determination unit is configured to perform rule responsibility determination based on the multi-source data to generate rule determination results; and to perform model determination based on the multi-source data to generate model determination results.

[0152] The analysis unit is configured to compare the rule-based determination result with the model-based determination result to output a difference result, wherein the model-based determination result is obtained by performing temporal analysis and spatial analysis on two neural network models respectively; and when the difference result is greater than or equal to a preset difference threshold, to perform interpretive analysis on the model-based determination result to output an analysis result, wherein the analysis result is the contribution of features in the multi-source data to the model-based determination result; and based on the analysis result, to determine a determination result, wherein the determination result is a rule-based determination result or a model-based determination result.

[0153] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. An explainable accident liability determination method characterized by comprising: The method comprises: acquiring multi-source data of an accident; based on the multi-source data, performing rule responsibility determination to generate a rule determination result, wherein the multi-source data comprises at least one of driving assistance event data, braking system event data, vehicle video data, bus data, and positioning data; based on the multi-source data, performing model determination to generate a model determination result, wherein the model determination result is a result of responsibility allocation obtained by performing time series analysis and spatial analysis based on two neural network models, the multi-source data comprises time series data and image data, the time series data comprises one or more combinations of vehicle speed data, turn signal data, brake pedal depth data, steering wheel angle data, and positioning trajectory data, and the image data comprises video data of multiple vehicle cameras; comparing the rule determination result and the model determination result to output a difference result; when the difference result is greater than or equal to a preset difference threshold, performing SHAP interpretive analysis on the model determination result to output an analysis result, wherein the analysis result is a contribution degree of a feature in the multi-source data to the model determination result; based on the analysis result, determining a determination result, wherein the determination result is the rule determination result or the model determination result; the determination of the determination result based on the analysis result comprises: determining whether the analysis result meets a preset traffic rule and common sense determination standard; when the analysis result meets the standard, determining that the model determination result is the determination result; when the analysis result does not meet the standard, determining that the rule determination result is the determination result.

2. The explainable accident liability determination method of claim 1, wherein, the rule responsibility determination based on the multi-source data to generate the rule determination result comprises: processing the multi-source data using a rule engine; based on a preset traffic rule knowledge base, performing logical determination to generate a determination result; generating the rule determination result according to the determination result. 3.The explainable accident liability determination method of claim 1, wherein, the model determination based on the multi-source data to generate the model determination result comprises: acquiring a time series neural network and a convolutional neural network, wherein the time series neural network is a long short-term memory network; processing time series data in the multi-source data through the time series neural network to extract time series features; processing image data in the multi-source data through the convolutional neural network to extract spatial features; fusing the time series features and the spatial features to generate fused features; generating the model determination result based on the fused features.

4. The explainable accident liability determination method of claim 3, wherein, the processing of the time series data in the multi-source data through the time series neural network to extract time series features comprises: performing preprocessing on the vehicle speed data, the turn signal data, the brake pedal depth data, the steering wheel angle data, and the positioning trajectory data to generate preprocessed time series data; inputting the preprocessed time series data into the long short-term memory network to extract hidden states representing vehicle motion patterns and driving behavior features through the long short-term memory network; determining the hidden states as the time series features.

5. The interpretable accident liability determination method of claim 3, wherein, The processing of the image data in the multi-source data by the convolutional neural network comprises: extracting key frame images from the video data; performing pre-processing on the key frame images to generate pre-processed images; inputting the pre-processed images into the convolutional neural network to extract image features from the bottom layer to the top layer through multi-layer convolution and pooling operations of the convolutional neural network; mapping the image features of the top layer to semantic information representing spatial relationships of traffic scenes; analyzing the semantic information to generate spatial features representing vehicle position relationships, lane line structures and road environment information.

6. The interpretable accident liability determination method of claim 1, wherein, The performing of the explanatory analysis on the model determination result to output an analysis result comprises: extracting features in the multi-source data; calculating contribution degrees of the features to the model determination result by an explanatory analysis method based on game theory; generating the analysis result based on the contribution degrees.

7. The interpretable accident liability determination method of claim 1, wherein, The method further comprises: labeling cases with the difference result greater than or equal to the preset difference threshold as conflict cases; adding multi-source data of the conflict cases to a training data set to obtain an updated data set; training a time series neural network and a convolutional neural network through the updated data set.

8. The interpretable accident liability determination method of claim 1, wherein, The performing of the rule responsibility determination based on the multi-source data to generate a rule determination result comprises: obtaining a first parameter, a second parameter, a third parameter, a fourth parameter and a fifth parameter; wherein the first parameter is an early warning event type and a trigger time extracted from the driving assistance event data, the second parameter is a brake activation state and a brake intensity obtained from the brake system event data, the third parameter is a traffic participant behavior and a road sign state identified from the vehicle video data, the fourth parameter is a vehicle operating state parameter parsed from the bus data, and the fifth parameter is a vehicle motion trajectory and a relative position relationship calculated from the positioning data; matching the first parameter, the second parameter, the third parameter, the fourth parameter and the fifth parameter with preset traffic rule conditions to generate a matching result; generating a rule determination result according to the matching result.

9. An explainable accident liability determination system, characterized by, comprise: an obtaining unit configured to obtain multi-source data of an accident; a result determination unit configured to perform rule responsibility determination based on the multi-source data to generate a rule determination result, wherein the multi-source data comprises at least one of the following: driving assistance event data, brake system event data, vehicle video data, bus data and positioning data when the rule responsibility determination is performed; and perform model determination based on the multi-source data to generate a model determination result, wherein the multi-source data comprises time series data and image data when the model determination is performed, the time series data comprises one or more combinations of vehicle speed data, turn signal data, brake pedal depth data, steering wheel angle data and positioning trajectory data, and the image data comprises video data of multiple vehicle cameras. The analysis unit is configured to compare the rule determination result with a model determination result to output a difference result, the model determination result being a result about responsibility allocation obtained after performing time series analysis and spatial analysis based on two neural network models respectively; and when the difference result is greater than or equal to a preset difference threshold, performing SHAP interpretive analysis on the model determination result to output an analysis result, the analysis result being a contribution degree of a feature in the multi-source data to the model determination result; and determining a determination result based on the analysis result, the determination result being the rule determination result or the model determination result; The determining the determination result based on the analysis result comprises: determining whether the analysis result meets a preset traffic rule and common sense determination standard; in the case of meeting, determining that the model determination result is the determination result; in the case of not meeting, determining that the rule determination result is the determination result.

Images