A large model-based traffic event analysis method

Abstract

The application discloses a traffic event analysis method based on a large model, and relates to the technical field of intelligent traffic video analysis, which comprises segmenting traffic monitoring video streams, using a pre-trained multi-modal large model to deeply analyze the semantic content of video clips, identifying potential and non-explicit abnormal traffic behaviors and complex interaction modes. Based on the identification results, a structured traffic event causal graph is constructed, and the graph is logically verified and information is completed through a pre-set traffic event knowledge base to form a standardized analysis basis. By subgraph mining of the graph, the key path and core node of event propagation are accurately identified. Based on the above analysis, an event report is generated and development is deduced, and an early warning and control scheme is output. The method can realize deep understanding of complex traffic situation and early discovery of implicit risks, and improve the explainability of event analysis and the accuracy of prediction through causal reasoning.

Description

Technical Field

[0001] This invention belongs to the field of intelligent traffic video analysis technology, specifically a traffic event analysis method based on a large model. Background Technology

[0002] Current traffic video event analysis primarily relies on specialized computer vision algorithms with predefined rules or detection models trained for specific events. These techniques typically target the identification of clear, apparent anomalies such as traffic accidents, traffic violations, and pedestrian intrusions. Their working principle is based on training models with large amounts of labeled data to identify visual patterns or behavioral features; essentially, it involves the classification and detection of known event types. While these methods are highly efficient for well-defined, distinctive target events, their perception and understanding capabilities are limited to predefined categories and rules, making them unable to handle the complex and ever-changing interactive behaviors and undefined, implicit risk patterns in traffic scenarios.

[0003] At the level of event correlation and trend analysis, existing solutions mostly remain at the level of simple statistics, ranking, or time-series-based prediction of detected events. They typically treat events as independent point data, fitting their probability of occurrence and trends through statistical regularities or machine learning models. This approach lacks structured modeling of the inherent causal logic and temporal relationships between events, leading to fragmented conclusions in event analysis that struggle to explain "why events occur" and "how events evolve." Predictions are often probabilistic outputs, lacking interpretable deductions based on causal chains, leaving subsequent control decisions without solid and transparent logical support. An analytical method is needed that can deeply understand complex traffic semantics and perform structured causal reasoning. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a traffic incident analysis method based on a large model, including: The raw video stream data of the traffic monitoring system is acquired and processed to obtain video analysis segments. The pre-trained multimodal large model is used to identify potential abnormal traffic patterns in the video analysis segment; Based on the identified potential abnormal traffic patterns, a traffic event causal graph is constructed, where nodes in the traffic event causal graph represent traffic elements or event states, and edges represent causal or temporal relationships. The traffic event causal graph is verified and completed using a pre-built traffic event knowledge base to generate a standardized traffic event graph. Subgraph mining is performed on the standardized traffic incident graph to identify key event propagation paths and core impact nodes; Based on the propagation path and core impact nodes of the key events, a traffic event analysis report is generated. The traffic event analysis report is then input into a traffic event prediction model to deduce the subsequent development trend of the traffic event. Based on the aforementioned subsequent development trends, traffic incident early warning signals and traffic control recommendations are generated.

[0005] Furthermore, the step of acquiring the raw video stream data from the traffic monitoring system and using a pre-trained multimodal large model to identify potential abnormal traffic patterns in the video analysis segments includes: The original video stream data is spatiotemporally sliced ​​to generate multiple video analysis segments; The video analysis segment is input into a pre-trained multimodal large model for video semantic understanding processing. Visual features and spatiotemporal correlation features of traffic elements in the video analysis segment are extracted to generate video semantic understanding results. Based on the video semantic understanding results, potential abnormal traffic patterns in the video analysis segment are identified. These potential abnormal traffic patterns include abnormal vehicle trajectory patterns, abnormal traffic participant interaction patterns, and abrupt traffic flow state patterns. The process of performing spatiotemporal slicing on the original video stream data to generate multiple video analysis segments includes: Detect scene switching points and keyframe sequences in the raw video stream data. The scene switching points include camera viewpoint switching and significant changes in traffic scenes. Based on the scene switching point, the original video stream data is divided into multiple consecutive primary video segments; Within each primary video segment, sliding window cutting is performed according to the preset time window length and sliding step size to generate intermediate video units of fixed duration; For each of the intermediate video units, key target detection is performed to identify moving vehicles, pedestrians, and traffic facilities in the video frames; Based on the key target detection results, the minimum bounding rectangle region containing all key targets is cropped out, and the minimum bounding rectangle region is subjected to resolution normalization processing to generate the final video analysis segment.

[0006] Furthermore, the step of inputting the video analysis segment into a pre-trained multimodal large model for video semantic understanding processing includes: The video analysis segment is split into image frame sequences and audio waveform data; The image frame sequence is input into the visual encoder of the multimodal large model to extract multi-level visual features, including pixel-level features, target-level features and scene-level features. The audio waveform data is input into the audio encoder of the multimodal large model to extract audio features, which include environmental sound features and abnormal sound features. The multi-level visual features and audio features are temporally aligned and fused to generate multimodal fusion features; The multimodal fusion features are input into the sequence understanding module of the multimodal large model. The spatiotemporal dependencies within the video segments are modeled through a self-attention mechanism, and the video semantic understanding results are output. The video semantic understanding results describe the traffic scene in the form of structured data.

[0007] Furthermore, the step of identifying potential abnormal traffic patterns in the video analysis segment based on the video semantic understanding results includes: The motion trajectory, speed changes, and direction information of traffic elements are extracted from the semantic understanding results of the video. The motion trajectory is matched with a preset normal traffic behavior template library to calculate the trajectory deviation, which includes lateral deviation, longitudinal deviation and speed deviation. The distance, relative speed, and interaction behavior between different traffic participants were extracted from the semantic understanding results of the video. The distance, relative speed, and interactive behavior are compared with a preset safe interaction rule base to detect whether there are conflicts, collision risks, or illegal interactions. Macro-level traffic flow parameters are extracted from the video semantic understanding results. These macro-level traffic flow parameters include traffic volume, average speed, and lane occupancy. Monitor the abrupt changes in the macroscopic traffic flow parameters over time to identify the formation, dissipation, or abnormal aggregation of traffic flow congestion; Based on the trajectory deviation, detection results, and abrupt change points, the potential abnormal traffic patterns are determined and labeled.

[0008] Furthermore, the step of constructing a causal graph of traffic events based on the identified potential abnormal traffic patterns includes: Each potential abnormal traffic pattern is instantiated as an event node, and the event node is labeled with the event type, occurrence time, spatial location, and severity. Analyze the temporal order and spatial proximity among the different event nodes, and establish temporal and spatial connection edges; Based on traffic rules and physical laws, infer the causal logic between the event nodes and establish causal connection edges. The causal logic includes triggering relationships, aggravating relationships, or inhibiting effects. Assign a confidence weight and an influence strength value to each connecting edge; By combining all the event nodes with connecting edges, an initial causal graph of the traffic events is constructed.

[0009] Furthermore, the step of verifying and completing the causal graph of traffic events using a pre-built traffic event knowledge base includes: The event nodes and connecting edges in the initial traffic event causal graph are matched with historical event cases in the traffic event knowledge base using graph structure similarity. Based on the matching results, relevant event evolution patterns and supplementary causal chains are retrieved from the traffic event knowledge base; Using the retrieved event evolution patterns and supplemented causal chains, missing nodes or edges in the initial traffic event causal graph are filled in, and low-confidence connection edges are corrected or deleted. The graph reasoning engine is invoked to deduce the existing hidden event nodes or indirect causal paths based on the completed graph structure, and these are then added to the graph. Output the standardized traffic event map after verification, completion, and inference optimization.

[0010] Furthermore, the subgraph mining process performed on the standardized traffic event map includes: In the standardized traffic event map, an event influence propagation model is defined, and the centrality index of each event node is calculated. The centrality index includes degree centrality, proximity centrality, and betweenness centrality. Based on the centrality index, the core influencing nodes are selected; Using the core impact nodes as the starting or ending points, a path search is performed in the standardized traffic event map to find all paths connecting two or more core impact nodes. Assess the cumulative impact intensity and propagation probability of each path, and select the paths with high impact intensity and high propagation probability as the propagation paths of the key events; Extract the connected subgraph containing the core influencing nodes and the propagation path of the key events, and use it as the core substructure for event analysis.

[0011] Furthermore, the generation of a traffic incident analysis report based on the key event propagation path and core impact nodes includes: The traffic incident analysis report includes the incident type, evolution process, scope of impact, and related elements; The propagation path of the key events is semantically described to generate text of the event evolution process. The core influencing nodes are attributed and their core triggering factors and main carriers are summarized. By combining the text of the event evolution process with the attributes of the core influencing nodes, the event type and the scope of influence are determined; Extract all traffic elements related to the event and their status changes from the standardized traffic event map to form the list of related elements; The event type, evolution process, scope of impact, and list of related elements are organized in a structured manner according to a preset report template to generate the traffic event analysis report.

[0012] Furthermore, the step of inputting the traffic incident analysis report into the traffic incident prediction model to deduce the subsequent development trend of the traffic incident includes: The traffic incident analysis report is analyzed to extract the static and dynamic features of the current incident. The static features include the incident type and location, and the dynamic features include the evolution speed and the trend of impact diffusion. The static and dynamic features are input into the traffic event prediction model, which is built based on recurrent neural networks and graph neural networks to model the temporal evolution and spatial diffusion of event states. Run the traffic incident prediction model to simulate changes in the incident status, expansion or contraction of the impact range, and potential secondary events at multiple future time steps; Output multiple development trend scenarios and assign an occurrence probability to each development trend scenario.

[0013] Furthermore, the traffic event prediction model is constructed through the following steps: Obtain a historical traffic incident case database, which includes historical traffic incident analysis reports and their corresponding actual subsequent development sequences; The features in the historical traffic incident analysis report are used as input, and the actual subsequent development sequence of the incident is used as training labels to form a training sample set. The initial network of the model is constructed using recurrent neural networks and graph neural networks as the basic architecture; The initial network of the model is trained under supervision using the training sample set. The network parameters are adjusted by optimizing the loss function until the error between the model prediction result and the actual development sequence converges to a preset range, thus obtaining the trained traffic event prediction model.

[0014] Compared with the prior art, the beneficial effects of the present invention are: Employing a pre-trained multimodal large-scale model as the core perception engine, this approach directly performs end-to-end analysis and understanding of video stream content. This large-scale model possesses powerful cross-modal information fusion and common-sense reasoning capabilities, enabling it to directly parse the dynamic interactions, behavioral intentions, and subtle deviations from smooth traffic flow patterns between traffic participants from the video's pixel sequence. This technological approach overcomes the limitations of traditional methods that rely on large amounts of specific event-labeled data to train dedicated models, achieving a generalized understanding of the deep semantics of video content. This allows for the identification of potential abnormal patterns that are not predefined but actually affect traffic efficiency and safety, representing a breakthrough from "explicit event detection" to "implicit situational awareness," providing a new information dimension for early warning of traffic risks.

[0015] By constructing a structured causal graph of traffic events and using a pre-built traffic event knowledge base for validation and completion, unstructured video semantic information is transformed into a graph network containing nodes and causal edges. This process formalizes the logical relationships between events and traffic elements. Subgraph mining and critical path analysis of this standardized graph clearly identify the core causal chains and key influencing factors leading to specific consequences. The traffic event prediction model uses this causal graph as input for deduction, and its prediction behavior is based on clear causal assumptions and logical relationships. Event analysis is elevated from a discrete list of labels to a networked analysis with complete logical chains and interpretable structures. The prediction results not only tell "what might happen," but also reveal "why it happened" and "where the key intervention points are," making the generated control recommendations more targeted and logically persuasive, and shifting the decision-making process from being based on statistical correlation to being based on causal correlation. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of the traffic incident analysis method based on a large model as described in this invention. Figure 2 A flowchart for video semantic understanding processing; Figure 3 A bar chart showing the graph structure similarity matching results during the verification and completion phase of the traffic incident graph. Figure 4 A bar chart for evaluating the propagation paths of key events during the subgraph mining phase of the causal graph of traffic incidents; Figure 5 This is a training convergence curve of the traffic incident prediction model. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1 The system acquires raw video stream data from a traffic monitoring system and processes it to obtain video analysis segments. A pre-trained multimodal large model is used to identify potential abnormal traffic patterns within these segments. Based on these patterns, a traffic event causal graph is constructed, where nodes represent traffic elements or event states, and edges represent causal or temporal relationships. A pre-built traffic event knowledge base is used to validate and complete the causal graph, generating a standardized graph. Subgraph mining is performed on the standardized graph to identify key event propagation paths and core impact nodes. Based on these paths, a traffic event analysis report is generated and input into a traffic event prediction model to predict future trends. Based on these trends, traffic event warning signals and traffic control recommendations are generated.

[0019] In one embodiment of the present invention, see [reference] Figure 2 The system acquires raw video stream data from intersections on main urban roads covered by a traffic monitoring system. This raw video stream data contains continuous sequences of image frames and synchronized audio streams. The raw video stream data undergoes spatiotemporal slicing, detecting scene transition points and keyframe sequences during the process. Exemplary scene transition points include perspective changes caused by traffic monitoring cameras' preset positions and significant changes in the traffic scene caused by traffic accidents, such as vehicles abruptly stopping in the center of an intersection. Based on the detected scene transition points, the several-hour-long raw video stream data is divided into multiple continuous primary video segments, each corresponding to a relatively stable monitoring perspective or traffic scene. Within a five-minute primary video segment, a sliding window segmentation operation is performed based on a preset two-minute time window length and a thirty-second sliding step, generating multiple fixed-duration intermediate video units. For each two-minute intermediate video unit, a key target detection algorithm is run to identify all moving vehicles, pedestrians, and traffic facilities such as traffic lights and lane markings in the video frame sequence. Based on the target bounding box sequence output by the key target detection algorithm, the smallest bounding rectangle region that can contain the temporal occurrence positions of all key targets is calculated and cropped. Resolution normalization processing is performed on the smallest bounding rectangle region, uniformly scaling it to 1920x1080 pixels, thereby generating the final video analysis segment used for analysis.

[0020] In some embodiments, video analysis segments are input into a pre-trained multimodal large-scale model for video semantic understanding processing. The processing decomposes the video analysis segments into a sequence of image frames extracted at 30 frames per second, and separate audio waveform data. The image frame sequence is fed into the visual encoder of the multimodal large-scale model, which extracts multi-level visual features based on a hierarchical convolutional neural network structure. These multi-level visual features include pixel-level features describing basic textures and edges, target-level features representing independent entities such as vehicles and pedestrians, and scene-level features summarizing the semantics of the entire intersection scene. The audio waveform data is fed into the audio encoder of the multimodal large-scale model, which extracts audio features based on Mel-spectrum mapping and convolutional networks. These audio features include environmental sound features reflecting background traffic noise, and abnormal sound features representing sudden braking, collisions, or horn honking. Through a temporal attention mechanism, the multi-level visual and audio features are aligned and fused in the temporal dimension to generate a unified multimodal fusion feature that integrates visual and auditory information. The multimodal fusion features are then fed into the sequence understanding module of the multimodal large model. The sequence understanding module is constructed using a Transformer architecture and models the spatiotemporal dependencies between frames and between objects within a video segment through a self-attention mechanism. The video semantic understanding results output by the sequence understanding module describe the traffic scene in the form of structured JSON data. The data structure includes the type, location, speed, and trajectory of each traffic participant, as well as the spatial relationships between participants and the overall state label of the scene.

[0021] Understandably, based on video semantic understanding results, potential abnormal traffic patterns in video analysis segments are identified. The identification process parses the motion trajectory coordinate sequence, speed change curve, and driving direction angle of each vehicle from the structured data of the video semantic understanding results. The parsed motion trajectory is then matched against a pre-defined normal traffic behavior template library, which defines trajectory templates such as normal driving within a lane and standard turns. The matching process calculates the trajectory deviation between the actual trajectory and the corresponding template trajectory. Trajectory deviation consists of three components: lateral deviation, longitudinal deviation, and speed deviation. Lateral deviation measures the degree to which the vehicle deviates from the lane centerline, longitudinal deviation measures abnormal acceleration or deceleration before the stop line, and speed deviation measures the difference between the vehicle's speed and the road speed limit or average traffic flow. Simultaneously, the distance matrix, relative speed vector, and interaction behavior labels, such as following, lane changing, and conflict, are parsed from the video semantic understanding results. The parsed distance, relative speed, and interaction behavior are compared with a pre-defined safety interaction rule library, which defines threshold rules such as safe following distance and safe lane changing interval. The comparison process detects the presence of illegal interaction tags such as distances below a safety threshold, excessively high relative speeds, or "conflicts." Furthermore, macroscopic traffic flow parameters are aggregated and extracted from the video semantic understanding results. These parameters include the traffic flow rate per minute across the cross-section, the average speed of all vehicles, and the lane occupancy rate (the proportion of time a lane is occupied by vehicles). Abrupt changes in macroscopic traffic flow parameters over time are monitored, and a sliding window statistical test method is used to identify points where traffic flow abruptly changes from a smooth to a congested state, points where congestion dissipates, or points of abnormal vehicle aggregation. The results of trajectory deviation calculations, illegal interaction detection, and traffic flow parameter abrupt change identification are combined and judged through a pre-defined decision logic. Finally, the video analysis segment is labeled with the presence of "abnormal vehicle trajectory patterns," "abnormal traffic participant interaction patterns," or "traffic flow state abrupt change patterns," and their specific subcategories.

[0022] Optionally, when calculating the overall trajectory deviation to aid in the judgment, the following formula can be used for quantitative evaluation: Where: characters in the formula Indicates the overall trajectory deviation, character This represents the calculated lateral deviation, in characters. This represents the calculated longitudinal deviation, represented by the character. This represents the calculated speed deviation, represented by the character. ,character ,character These are pre-set weighting coefficients used to adjust the importance of different deviation components in the overall evaluation.

[0023] In some embodiments, the recognition process of a video analysis segment generates specific data. For example, the calculated lateral deviation of a car's trajectory is 0.85, the calculated longitudinal deviation is 0.20, and the calculated speed deviation is 1.50. With weighting coefficients α = 0.4, β = 0.3, and γ = 0.3, the calculated comprehensive trajectory deviation is 0.91. Simultaneously, the safety interaction rule base detects that the car is too close to a truck in the adjacent lane, with a large relative speed difference, marking it as a high-risk interaction. Macroscopic traffic flow parameter monitoring shows that within 30 seconds of the car's abnormal behavior, the average speed in its lane decreases from 60 km / h to 20 km / h, and the lane occupancy increases from 15% to 70%, identifying a traffic flow state abrupt change. Combining this data, it is determined that the video analysis segment simultaneously exhibits both an "abnormal vehicle trajectory pattern" and a "normal traffic participant interaction pattern," and has triggered a "traffic flow state abrupt change pattern."

[0024] In practice, through the above process, the raw video stream data is transformed into a series of structured semantic units labeled with potential abnormal traffic patterns. These structured semantic units accurately describe the occurrence time, spatial location, involved entities, and quantitative indicators of abnormal events, providing clear input for the subsequent construction of traffic event causal maps. The multimodal large-scale model video semantic understanding processing and multi-level abnormal pattern recognition rules together complete the transformation from raw pixels to high-level traffic semantic events.

[0025] In one embodiment of the present invention, a traffic event causal graph is constructed based on identified potential abnormal traffic patterns. The construction process begins by instantiating each identified potential abnormal traffic pattern as an independent event node, with detailed annotations of the event node's attributes. For example, a "vehicle trajectory abnormal pattern" is instantiated as an event node, labeled with the event type "motor vehicle sudden braking," the occurrence time "2025-10-26 08:15:32," the spatial location "latitude and longitude coordinates (X1, Y1)," and the severity "moderate." Another "traffic participant interaction abnormal pattern" is instantiated as another event node, labeled with the event type "motor vehicle-non-motor vehicle conflict risk," the occurrence time "2025-10-26 08:15:35," the spatial location "latitude and longitude coordinates (X2, Y2)," and the severity "altitude."

[0026] In some embodiments, through deep semantic understanding of multimodal large models, the identified potential abnormal traffic patterns are not limited to vehicle trajectories, participant interactions, and macroscopic traffic flow anomalies. For example, the model can further identify specific events such as 'large items of garbage' scattered due to unsecured vehicle loads, 'illegal parking' caused by vehicles staying in non-parking areas for extended periods, 'smoke and fire' caused by vehicle fires or roadside construction, 'accidents' such as collisions or rollovers between vehicles, 'road flooding' caused by rainfall or pipe ruptures, 'facility damage' such as traffic light malfunctions or fallen guardrails, and 'potholes' caused by road surface structural damage. These events are identified by parsing target-level and scene-level visual features from the image frame sequence of video analysis segments through the aforementioned steps, or by combining abnormal sound features from audio waveform data, and then comparing them with a pre-set library of various abnormal event behavior templates and a safety rule library. Instances of these newly added event types are also instantiated as event nodes and labeled with event type, time, location, and other information for subsequent causal graph construction.

[0027] In some embodiments, the temporal order and spatial proximity relationships between different event nodes are analyzed to establish temporal and spatial connection edges. The analysis process calculates the timestamp difference between one event node and another. A timestamp difference of 3 seconds meets the preset temporal association threshold of "within 10 seconds," therefore, a temporal connection edge is established between the event node and another event node, with the edge pointing from the earlier-occurring event node to the later-occurring event node. Simultaneously, the Euclidean distance between the event node and another event node is calculated. A spatial Euclidean distance of 15 meters meets the preset spatial proximity threshold of "within 50 meters," therefore, a spatial connection edge is established between the event node and another event node. This spatial connection edge is undirected.

[0028] It is understandable that, based on traffic rules and physical laws, the causal logic between event nodes is inferred, and causal connections are established. The inference process is based on the type of event node, its spatiotemporal relationship, and common sense about physics. The location of event node "sudden braking of a motor vehicle" is upstream of another event node "risk of conflict between motor vehicles and non-motor vehicles," and it occurs 3 seconds earlier than the other event node. According to the laws of traffic movement, the sudden braking of the vehicle in front may cause traffic participants behind to react too late or take evasive action, thereby increasing the risk of conflict. Therefore, it is inferred that there is a triggering relationship between event node and another event node. A causal connection is established from event node to another event node, and the type of causal connection is "triggering." If multiple event nodes simultaneously exacerbate the severity of a certain event, causal connections with an aggravating relationship may be established.

[0029] Optionally, each connection edge can be assigned a confidence weight and an influence strength value. The confidence weight reflects the certainty of the existence of a causal or temporal relationship, while the influence strength value reflects the strength of the source event node's influence on the target event node. For example, the confidence weight of the causal connection edge in the above-mentioned "trigger" relationship can be calculated based on the time interval and spatial distance; the shorter the time interval and the closer the spatial distance, the higher the confidence weight. The influence strength value of this causal connection edge can be quantified according to the severity (moderate) of the event node "sudden braking of a motor vehicle" and the event type association rules. One formula for calculating the influence strength value is: Where: characters This represents the calculated influence strength value, character. This represents the baseline impact coefficient determined based on event type association rules, in characters. The character represents the quantified severity value of the source event node (sudden braking of a motor vehicle). The maximum value for severity quantization is represented by the character. Indicates the time decay coefficient, character This indicates the time interval between the source event node and the target event node.

[0030] In practice, all event nodes and connecting edges are combined to construct an initial causal graph of traffic events. The combination process uses each instantiated and labeled event node as a vertex of the graph, and each established temporal, spatial, and causal connecting edge as a directed or undirected edge. Connecting edges are associated with confidence weights and influence strength values. These elements are stored and linked using a graph database or graph structure data model to form a directed graph with attributes and weights; this directed graph is the initial causal graph of traffic events. The initial causal graph of traffic events, in visual or data form, explicitly displays the spatiotemporal relationships and inferred causal logic between multiple abnormal traffic events, providing a core data structure foundation for subsequent graph verification and completion.

[0031] In one embodiment of the present invention, a pre-built traffic event knowledge base is used to verify and complete an initial traffic event causal graph, generating a standardized traffic event graph. The verification and completion process begins by matching the event nodes and connecting edges in the initial traffic event causal graph with historical event cases in the traffic event knowledge base using graph structure similarity matching. The graph structure similarity matching process calculates the topological similarity and attribute similarity between the subgraphs of the initial traffic event causal graph and the historical event case graph. For example, the initial traffic event causal graph contains a "motor vehicle sudden braking" event node connected to a "motor vehicle-non-motor vehicle conflict risk" event node. Subgraph structures with similar node types and connecting edge types are retrieved from the historical event cases in the traffic event knowledge base. The matching process may find a historical case whose graph shows that the "motor vehicle yielding to pedestrian" event is connected to the "rear vehicle deceleration" event through a "triggering" relationship, and then connected to the "intersection local congestion" event through a "triggering" relationship. This historical case graph has local structural similarity to the initial traffic event causal graph.

[0032] In some embodiments, based on the matching results, relevant event evolution patterns and supplementary causal chains are retrieved from the traffic event knowledge base. For example, through the above matching, the event evolution pattern of "motor vehicle-vulnerable road user conflict" is retrieved. This event evolution pattern describes a typical pattern from "emergency braking of a motor vehicle" to "increased risk of traffic conflict," and then to "chain reaction of subsequent vehicles." Simultaneously, supplementary causal chains are retrieved, indicating that if the "increased risk of traffic conflict" event occurs at an intersection, it may further "trigger" an event not included in the initial traffic event causal map: "decreased traffic efficiency during the traffic signal cycle."

[0033] It is understandable that the retrieved event evolution patterns and supplementary causal chains are used to complete missing nodes or edges in the initial traffic event causal graph, and low-confidence connections are corrected or deleted. The completion process adds a new "decrease in traffic efficiency during a traffic signal cycle" event node to the initial traffic event causal graph based on the retrieved supplementary causal chains, and establishes a causal connection edge from the existing "motor vehicle-non-motor vehicle conflict risk" event node to this new event node. The causal connection edge type is "triggered". The correction process targets connections that may exist in the initial traffic event causal graph but whose confidence weight is below a preset threshold. For example, a causal connection edge based on fuzzy spatiotemporal relationships connecting "distant vehicle lane change" and "near-end sudden braking" is deleted from the graph due to its low confidence weight and lack of supporting evidence in the traffic event knowledge base.

[0034] Optionally, the graph structure similarity matching process can be evaluated using a quantitative similarity calculation formula: Where: characters This represents the calculated graph structure similarity score, where characters... and characters Let the weight coefficients of topological similarity and attribute similarity be respectively, and satisfy the following conditions: ,character Represents the topological similarity value calculated based on graph isomorphism or edit distance, character This represents the attribute similarity value calculated based on attributes such as event node type and severity.

[0035] In practical implementation, the graph inference engine is invoked. Based on the completed graph structure, it deduces existing hidden event nodes or indirect causal paths and adds them to the graph. The graph inference engine operates based on rule-based reasoning or graph neural network reasoning methods. For example, the graph inference engine applies a predefined domain rule: "If event A (sudden braking of a motor vehicle) triggers event B (conflict risk), and event B occurs in an intersection area, then there may be an unobserved event C (pedestrian or non-motorized vehicle running a red light) as a potential cause of event A." According to this rule, the graph inference engine deduces a new "non-motorized vehicle running a red light" event node as a hidden event node and establishes a causal connection edge from the new "non-motorized vehicle running a red light" event node to the "sudden braking of a motor vehicle" event node. Simultaneously, the graph inference engine may discover that the "sudden braking of a motor vehicle" event node indirectly affects the "decreased traffic efficiency" event node through the "conflict risk" event node, thereby deduceing an indirect causal path and labeling the explicit relationships of this indirect causal path.

[0036] In some embodiments, a standardized traffic event graph is output, after verification, completion, and inference optimization. Compared to the initial traffic event causal graph, the standardized traffic event graph includes event nodes and causal connections completed from the traffic event knowledge base, removes low-confidence connections, and adds implicit event nodes and indirect relationships derived by the graph inference engine. The standardized traffic event graph is stored using a unified and standardized graph data structure, where each element has a definite source label. The standardized traffic event graph provides a more complete and reliable description of the causal and logical relationships of traffic events.

[0037] See Figure 3This is a bar chart showing the graph structure similarity matching results used in the verification and completion phase of a traffic event graph. It primarily displays the topological similarity, attribute similarity, and weighted comprehensive similarity scores of five historical cases, serving as a quantitative basis for structural matching between the traffic event causal graph and historical cases. This chart can directly support the verification and completion of the traffic event causal graph. Cases with high comprehensive similarity (such as Case 3) should be prioritized, as their historical event evolution patterns and supplementary causal chains can be directly used to complete the current graph. For cases with high attribute similarity but low topological similarity (such as Case 4), their event node attributes can be extracted as supplementary evidence, but their causal chain structure should be used with caution. Based on the weighting rules, topological similarity has a greater impact on the comprehensive score, indicating that in traffic event graph matching, the structural rationality of the causal chain is more critical than the attribute similarity of individual events.

[0038] In one embodiment of the present invention, subgraph mining is performed on a standardized traffic event graph to identify key event propagation paths and core influencing nodes. The process defines an event influence propagation model within the standardized traffic event graph. This model is constructed based on a random walk or independent cascade model to simulate the transmission of influence between event nodes. The centrality index of each event node in the standardized traffic event graph is calculated. The centrality index includes degree centrality, proximity centrality, and betweenness centrality. Degree centrality counts the number of edges directly connected to an event node; proximity centrality calculates the reciprocal of the sum of the shortest path lengths from an event node to all other event nodes in the graph; and betweenness centrality counts the number of times an event node lies on the shortest path between any two nodes in the graph.

[0039] In some embodiments, core impact nodes are selected based on centrality indicators. The selection process normalizes the different centrality indicator values ​​for each event node and performs a weighted summation according to predefined weights to obtain a comprehensive core impact node selection value. For example, an event node named "Motor Vehicle-Non-Motor Vehicle Conflict Risk" has a degree centrality of 5, a proximity centrality of 0.85, and a betweenness centrality of 12. Through weighted calculation, the core impact node selection value of this event node is significantly higher than that of other event nodes, and therefore it is determined to be a core impact node. Refer to Table 1, which shows example data of the centrality indicators and core impact node selection values ​​for some event nodes.

[0040] Table 1: Event Node Centrality Indicators and Core Influence Node Screening Values It is understandable that, starting from or ending at a core impact node, a path search is performed within a standardized traffic event graph to find all paths connecting two or more core impact nodes. The path search process employs a depth-first or breadth-first graph traversal algorithm. Starting from the core impact node "E002" (motor vehicle-non-motor vehicle conflict risk) in Table 1, the search proceeds along the connecting edges within the standardized traffic event graph. The search process identifies paths connecting core impact nodes "E002" and "E003" (chain reaction deceleration of vehicles behind), as well as paths connecting core impact nodes "E002" and "E001" (sudden braking of motor vehicles). It may also identify longer paths that connect multiple core impact nodes via other non-core impact nodes.

[0041] Optionally, the cumulative impact strength and propagation probability of each path are evaluated, and paths with high impact strength and high propagation probability are selected as key event propagation paths. The cumulative impact strength is calculated by aggregating the impact strength values ​​of all connecting edges along the path, such as by summing or multiplying. The propagation probability is calculated using an event influence propagation model, reflecting the likelihood of the event evolving along that path. One formula for calculating the cumulative impact strength of a path is: Where: characters This represents the calculated cumulative impact strength of the path, represented by the character. This indicates the total number of connecting edges contained in the path, character. Indicates the first path The influence strength value attached to each connecting edge. Calculate the cumulative influence strength of each path obtained from the search. Based on the propagation probability, a threshold is set, and paths with a cumulative impact intensity higher than the threshold of 0.8 and a propagation probability higher than the threshold of 0.7 are selected as critical event propagation paths.

[0042] In practice, a connected subgraph containing core impact nodes and key event propagation paths is extracted as the core substructure for event analysis. The extraction process is based on the selected set of core impact nodes, traversing all event nodes and connecting edges along the key event propagation paths. These elements are completely extracted from the standardized traffic event graph, forming a smaller, more tightly structured connected subgraph. This connected subgraph focuses on the most crucial event chains and causal relationships, constituting the direct object of subsequent in-depth analysis.

[0043] In some embodiments, a traffic incident analysis report is generated based on the propagation path of key events and core impact nodes. The generation process specifies that the traffic incident analysis report includes the event type, evolution process, scope of impact, and related elements. The propagation path of key events is semantically described, converting the sequence of event nodes on the path and their connections into natural language text to generate text on the event evolution process. For example, a key event propagation path is described as: "At 08:15:32, a sudden braking event of a motor vehicle occurred at location (X1,Y1); subsequently at 08:15:35, a motor vehicle-non-motor vehicle conflict risk event was triggered at location (X2,Y2) 15 meters downstream; this conflict risk further caused multiple vehicles behind to slow down in a chain reaction at 08:16:00." It is understandable that the core impact nodes are attributed to summarize the core triggering factors and main carriers of the event. For the core impact node "E002" (motor vehicle-non-motor vehicle conflict risk) in Table 1, its attributes are summarized as follows: the core triggering factor is "close-range high-speed interaction between motor vehicles and vulnerable road users," and the main carriers are "the involved motor vehicle and non-motor vehicle, and the lane in which they are located." Combining the event evolution text with the attribute summary of the core impact nodes, the event type is determined to be "a chain traffic conflict event triggered by emergency braking," and the scope of impact is determined to include "the lane where the incident occurred and an adjacent lane within a 50-meter section, lasting approximately 2 minutes."

[0044] In practice, all traffic elements and their state changes related to the event are extracted from a standardized traffic event map to form a list of associated elements. The extraction process traverses all event nodes involved in the core substructure of the event analysis, parsing the traffic elements recorded in the event node attributes. The list of associated elements includes specific elements such as "a gray sedan with license plate A12345 (speed decreased from 50km / h to 0km / h)", "a white SUV with license plate B67890 (speed decreased from 45km / h to 20km / h)", "a cyclist (passing through the conflict point)", and "the eastbound straight lane (occupancy increased from 15% to 70%)". The determined event type, the generated event evolution process text, the determined scope of impact, and the formed list of associated elements are structured and filled according to a pre-set report template with fixed fields, ultimately generating a complete traffic event analysis report document.

[0045] See Figure 4This is a bar chart used in the subgraph mining phase of a traffic incident causal graph to evaluate the propagation paths of critical events. It primarily displays the impact strength and propagation probability of three potential propagation paths, and uses a critical path threshold of 0.8 to filter out the truly critical event propagation paths. Path 1 and Path 2 are identified as critical event propagation paths requiring focused monitoring, providing core evidence for subsequent traffic incident analysis and early warning. Path 2 carries the highest risk and should be the primary target for traffic control, with priority given to deploying early warning and intervention measures. Based on this result, monitoring and response resources can be concentrated on critical paths, improving the efficiency and accuracy of traffic incident response. The evaluation results of the critical paths in the chart can be incorporated into the traffic incident knowledge base as core features of historical cases, providing data support for the graph verification and completion of similar events, and continuously optimizing the accuracy of event analysis.

[0046] In one embodiment of the present invention, a traffic incident analysis report is input into a traffic incident prediction model to extrapolate the subsequent development trend of the traffic incident. The extrapolation process begins by parsing the traffic incident analysis report, extracting the static and dynamic features of the current event from the structured fields of the report. Static features include specific categories parsed from the "Event Type" field of the traffic incident analysis report, such as "a chain traffic conflict incident triggered by emergency braking," and specific location descriptions parsed from the "Affected Area" field, such as "the eastern entrance of the intersection of XX Road and YY Road." Dynamic features include the evolution speed quantitatively analyzed from the "Evolution Process" text of the traffic incident analysis report, such as the growth rate of the number of affected vehicles per unit time, and the impact diffusion trend inferred based on changes in traffic flow parameters in the list of related elements, such as the direction and speed of congestion spreading upstream from the lane where the incident occurred.

[0047] In some embodiments, the extracted static and dynamic features are input into a traffic event prediction model, which is constructed based on recurrent neural networks (RNNs) and graph neural networks (GNNs). The RNN portion, such as a long short-term memory network (LSTM), models the temporal evolution of event states over time. The GNN portion, such as a graph convolutional network (GCN), models the spatial diffusion of event impacts across a spatial topology defined by the spatial relationships between event nodes in a standardized traffic event graph. The input layer of the traffic event prediction model receives vectors representing the static and dynamic features. These vectors are then jointly processed by the RNN and GNN layers to output a prediction of the future event state.

[0048] It is understandable that running a traffic event prediction model simulates changes in the event state, the expansion or contraction of its impact range, and potential secondary events over multiple future time steps. The simulation process uses the event state at the current time point as the initial state, and the traffic event prediction model uses the prediction output of the previous time step as part of the input for the next time step, iteratively extrapolating. For example, simulating in the next 5 minutes (time step t1), the event state evolves from "conflict risk" to "local congestion," and the impact range expands from "the affected lane and approximately 50 meters of adjacent lanes" to "100 meters upstream." Simulating in the next 10 minutes (time step t2), the event state may remain "local congestion" or, based on a certain probability, evolve into "secondary traffic accident," and the impact range may further expand or begin to contract. The traffic event prediction model outputs a series of possible event states and their corresponding spatial impact range changes for each simulated future time step.

[0049] Optionally, the traffic event prediction model outputs multiple development trend scenarios and assigns a probability of occurrence to each scenario. The model achieves multi-scenario prediction through its internal probability output layer or multiple Monte Carlo sampling. For example, for the next 10 minutes, the model might output three main development trend scenarios: Scenario A, "Congestion gradually dissipates," with a probability of 0.6; Scenario B, "Congestion continues and slightly worsens," with a probability of 0.3; and Scenario C, "Secondary rear-end collision occurs," with a probability of 0.1. Each scenario is accompanied by a detailed description of the event state and quantified impact parameters.

[0050] In its implementation, the traffic incident prediction model is constructed through the following steps: A historical traffic incident case library is obtained, containing historical traffic incident analysis reports and their corresponding subsequent actual development sequences. The historical traffic incident analysis reports serve as prototypes for the model's input features. The subsequent actual development sequences record the actual state changes, the evolution of the impact range, and whether secondary events occurred after the generation time of the historical traffic incident analysis reports. These subsequent actual development sequences serve as the target labels for model training.

[0051] In some embodiments, features from historical traffic incident analysis reports are used as input, and the subsequent actual development sequence of the incident is used as training labels to form a training sample set. The process of constructing the training sample set requires performing the same feature extraction operation on the historical traffic incident analysis reports as in real-time analysis, obtaining static and dynamic feature vectors. The static and dynamic feature vectors are concatenated to form the model input feature vector. The subsequent actual development sequence of the incident is encoded into a label vector in time-series form. Each pair (input feature vector, label vector) constitutes a training sample, and a large set of training samples constitutes the training sample set.

[0052] It can be understood that the initial network of the model is constructed using recurrent neural networks (RNNs) and graph neural networks (Graph NNNs) as the basic architecture. The specific configuration of the initial network includes a two-layer long short-term memory (LSTM) network for the RNN and a two-layer graph attention network for the Graph NNN. The outputs of the RNN and Graph NNN are fused in an intermediate layer, and the fused features are then mapped to an output space consistent with the label dimension through a fully connected layer. The initial network defines a learnable functional relationship from input features to predicted output.

[0053] In practice, supervised training of the initial network is performed using a training sample set, and the network parameters are adjusted by optimizing the loss function. The supervised training process uses mean squared error or cross-entropy, commonly used in time series prediction, as the loss function. The loss function calculates the difference between the initial network's predicted output for each training sample and the actual subsequent event label vector. Through backpropagation and gradient descent optimizer, the weight parameters of the long short-term memory network, graph attention network, and fully connected layers in the initial network are iteratively adjusted to continuously reduce the value of the loss function. Training continues until the error between the model's prediction and the actual event converges to a preset range; for example, the loss function value on the validation set no longer decreases significantly over multiple training cycles. At this point, the trained traffic event prediction model is obtained. The trained traffic event prediction model can generalize and extrapolate the subsequent development trend of similar traffic events based on newly input traffic event analysis reports.

[0054] See Figure 5This is a training convergence curve of a traffic event prediction model, showing the trend of the mean squared error of the training and validation sets as the number of training iterations changes. It is primarily used to evaluate the model's training effectiveness and generalization ability. The validation set loss continuously decreases and stabilizes, indicating that the traffic event prediction model has good generalization ability and can effectively extrapolate the development trend of events in real-world scenarios. The validation set loss no longer decreases significantly after 80 iterations, which can be used as the termination point for model training, avoiding resource waste and overfitting caused by overtraining. The final validation set loss (approximately 0.13) can serve as a quantitative benchmark for model performance, used for comparing the effects of subsequent iterations. The continuous decrease in training set loss and the stabilization of validation set loss prove that the model has not overfitted, its generalization ability is reliable, and it provides crucial risk verification evidence for model deployment.

[0055] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A traffic incident analysis method based on a large model, characterized in that, include: The raw video stream data of the traffic monitoring system is acquired and processed to obtain video analysis segments. The pre-trained multimodal large model is used to identify potential abnormal traffic patterns in the video analysis segment; Based on the identified potential abnormal traffic patterns, a traffic event causal graph is constructed, where nodes in the traffic event causal graph represent traffic elements or event states, and edges represent causal or temporal relationships. The traffic event causal graph is verified and completed using a pre-built traffic event knowledge base to generate a standardized traffic event graph. Subgraph mining is performed on the standardized traffic incident graph to identify key event propagation paths and core impact nodes; Based on the propagation path and core impact nodes of the aforementioned key events, a traffic incident analysis report is generated. This report is then input into a traffic incident prediction model to deduce the subsequent development trend of the traffic incident, including: The traffic incident analysis report is analyzed to extract the static and dynamic features of the current incident. The static features include the incident type and location, and the dynamic features include the evolution speed and the trend of impact diffusion. The static and dynamic features are input into the traffic event prediction model, which is built based on recurrent neural networks and graph neural networks to model the temporal evolution and spatial diffusion of event states. Run the traffic incident prediction model to simulate changes in the incident status, expansion or contraction of the impact range, and potential secondary events at multiple future time steps; Output multiple development trend scenarios and assign an occurrence probability to each development trend scenario; Based on the aforementioned subsequent development trends, traffic incident early warning signals and traffic control recommendations are generated.

2. The traffic incident analysis method based on a large model according to claim 1, characterized in that, The process of acquiring raw video stream data from a traffic monitoring system and using a pre-trained multimodal large model to identify potential abnormal traffic patterns in the video analysis segments includes: The original video stream data is spatiotemporally sliced ​​to generate multiple video analysis segments; The video analysis segment is input into a pre-trained multimodal large model for video semantic understanding processing. Visual features and spatiotemporal correlation features of traffic elements in the video analysis segment are extracted to generate video semantic understanding results. Based on the video semantic understanding results, potential abnormal traffic patterns in the video analysis segment are identified. These potential abnormal traffic patterns include abnormal vehicle trajectory patterns, abnormal traffic participant interaction patterns, and abrupt traffic flow state patterns. The process of performing spatiotemporal slicing on the original video stream data to generate multiple video analysis segments includes: Detect scene switching points and keyframe sequences in the raw video stream data. The scene switching points include camera viewpoint switching and significant changes in traffic scenes. Based on the scene switching point, the original video stream data is divided into multiple consecutive primary video segments; Within each primary video segment, sliding window cutting is performed according to the preset time window length and sliding step size to generate intermediate video units of fixed duration; For each of the intermediate video units, key target detection is performed to identify moving vehicles, pedestrians, and traffic facilities in the video frames; Based on the key target detection results, the minimum bounding rectangle region containing all key targets is cropped out, and the minimum bounding rectangle region is subjected to resolution normalization processing to generate the final video analysis segment.

3. The traffic incident analysis method based on a large model according to claim 2, characterized in that, The step of inputting the video analysis segment into a pre-trained multimodal large model for video semantic understanding processing includes: The video analysis segment is split into image frame sequences and audio waveform data; The image frame sequence is input into the visual encoder of the multimodal large model to extract multi-level visual features, including pixel-level features, target-level features and scene-level features. The audio waveform data is input into the audio encoder of the multimodal large model to extract audio features, which include environmental sound features and abnormal sound features. The multi-level visual features and audio features are temporally aligned and fused to generate multimodal fusion features; The multimodal fusion features are input into the sequence understanding module of the multimodal large model. The spatiotemporal dependencies within the video segments are modeled through a self-attention mechanism, and the video semantic understanding results are output. The video semantic understanding results describe the traffic scene in the form of structured data.

4. The traffic incident analysis method based on a large model according to claim 2, characterized in that, The step of identifying potential abnormal traffic patterns in the video analysis segment based on the video semantic understanding results includes: The motion trajectory, speed changes, and direction information of traffic elements are extracted from the semantic understanding results of the video. The motion trajectory is matched with a preset normal traffic behavior template library to calculate the trajectory deviation, which includes lateral deviation, longitudinal deviation and speed deviation. The distance, relative speed, and interaction behavior between different traffic participants were extracted from the semantic understanding results of the video. The distance, relative speed, and interactive behavior are compared with a preset safe interaction rule base to detect whether there are conflicts, collision risks, or illegal interactions. Macro-level traffic flow parameters are extracted from the video semantic understanding results. These macro-level traffic flow parameters include traffic volume, average speed, and lane occupancy. Monitor the abrupt changes in the macroscopic traffic flow parameters over time to identify the formation, dissipation, or abnormal aggregation of traffic flow congestion; Based on the trajectory deviation, detection results, and abrupt change points, the potential abnormal traffic patterns are determined and labeled.

5. The traffic incident analysis method based on a large model according to claim 1, characterized in that, The step of constructing a traffic event causal graph based on the identified potential abnormal traffic patterns includes: Each potential abnormal traffic pattern is instantiated as an event node, and the event node is labeled with the event type, occurrence time, spatial location, and severity. Analyze the temporal order and spatial proximity among the different event nodes, and establish temporal and spatial connection edges; Based on traffic rules and physical laws, infer the causal logic between the event nodes and establish causal connection edges. The causal logic includes triggering relationships, aggravating relationships, or inhibiting effects. Assign a confidence weight and an influence strength value to each connecting edge; By combining all the event nodes with connecting edges, an initial causal graph of the traffic events is constructed.

6. The traffic incident analysis method based on a large model according to claim 5, characterized in that, The process of verifying and completing the causal graph of traffic events using a pre-built traffic event knowledge base includes: The event nodes and connecting edges in the initial traffic event causal graph are matched with historical event cases in the traffic event knowledge base using graph structure similarity. Based on the matching results, relevant event evolution patterns and supplementary causal chains are retrieved from the traffic event knowledge base; Using the retrieved event evolution patterns and supplemented causal chains, missing nodes or edges in the initial traffic event causal graph are filled in, and low-confidence connection edges are corrected or deleted. The graph reasoning engine is invoked to deduce the existing hidden event nodes or indirect causal paths based on the completed graph structure, and these are then added to the graph. Output the standardized traffic event map after verification, completion, and inference optimization.

7. The traffic incident analysis method based on a large model according to claim 1, characterized in that, The subgraph mining process for the standardized traffic event map includes: In the standardized traffic event map, an event influence propagation model is defined, and the centrality index of each event node is calculated. The centrality index includes degree centrality, proximity centrality, and betweenness centrality. Based on the centrality index, the core influencing nodes are selected; Using the core impact nodes as the starting or ending points, a path search is performed in the standardized traffic event map to find all paths connecting two or more core impact nodes. Assess the cumulative impact intensity and propagation probability of each path, and select the paths with high impact intensity and high propagation probability as the propagation paths of the key events; Extract the connected subgraph containing the core influencing nodes and the propagation path of the key events, and use it as the core substructure for event analysis.

8. The traffic incident analysis method based on a large model according to claim 7, characterized in that, The traffic incident analysis report, generated based on the key event propagation path and core impact nodes, includes: The traffic incident analysis report includes the incident type, evolution process, scope of impact, and related elements; The propagation path of the key events is semantically described to generate text of the event evolution process. The core influencing nodes are attributed and their core triggering factors and main carriers are summarized. By combining the text of the event evolution process with the attributes of the core influencing nodes, the event type and the scope of influence are determined; Extract all traffic elements related to the event and their status changes from the standardized traffic event map to form the list of related elements; The event type, evolution process, scope of impact, and list of related elements are organized in a structured manner according to a preset report template to generate the traffic event analysis report.

9. The traffic incident analysis method based on a large model according to claim 1, characterized in that, The traffic incident prediction model is constructed through the following steps: Obtain a historical traffic incident case database, which includes historical traffic incident analysis reports and their corresponding actual subsequent development sequences; The features in the historical traffic incident analysis report are used as input, and the actual subsequent development sequence of the incident is used as training labels to form a training sample set. The initial network of the model is constructed using recurrent neural networks and graph neural networks as the basic architecture; The initial network of the model is trained under supervision using the training sample set. The network parameters are adjusted by optimizing the loss function until the error between the model prediction result and the actual development sequence converges to a preset range, thus obtaining the trained traffic event prediction model.

Images