Highway video event real-time pushing method based on edge node cooperation

Abstract

The application discloses a highway video event real-time pushing method based on edge node cooperation and relates to the technical field of intelligent traffic analysis. The method comprises the following steps: processing video streams by each edge node, generating local event segment data and its announcement message; performing first-stage matching based on space-time overlap and abstract information through broadcast announcement message between nodes, obtaining detailed event segment data for second-stage matching and time alignment after realizing lightweight cooperative discovery, fusing to generate cross-node event chain data and giving initial fusion credibility; for the period with insufficient credibility, obtaining compensation data from other nodes and updating the fusion credibility according to motion continuity; mapping the event chain data to a road space grid, calculating space-time occupation intensity and attenuation, generating a dynamic space occupation graph, calculating event impact metric values according to the fusion credibility, finally sorting to generate a pushing priority instruction and sending an event announcement. The application improves the real-time performance of event response and the traffic command efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic analysis technology, specifically to a method for real-time push of highway video events based on edge node collaboration. Background Technology

[0002] Intelligent push technology for highway video events based on edge node collaboration is an important development direction in the field of intelligent transportation. It aims to achieve rapid detection and accurate response to abnormal events such as traffic accidents and congestion through real-time perception and collaborative computing of distributed video analysis nodes. In recent years, with the advancement of edge computing and computer vision technologies, highway event processing solutions have gradually evolved from centralized processing to distributed, collaborative real-time perception.

[0003] Existing technologies mostly employ a method of local event detection at the camera and directly reporting the results to a central cloud platform. The central platform merges and deduplicates event reports from different nodes through simple temporal and spatial correlations, forming a preliminary event list and pushing it according to preset rules. These methods achieve a certain degree of automated event discovery, reduce reliance on manual inspections, and can provide tiered responses based on event severity, possessing a certain degree of real-time capability and practicality.

[0004] However, in real-world distributed highway monitoring scenarios, the field of view of a single camera is limited and susceptible to interference from factors such as obstruction, weather, and changes in lighting. This results in the "event fragments" reported by a single node often being partial and incomplete, with fluctuating detection reliability. Due to clock differences, frame rate variations, and misaligned viewpoints, the observations of the same continuous event by different nodes are fragmented and asynchronous in both time and space. Existing technologies struggle to effectively address this "information fragmentation" problem caused by asynchronous and unreliable observations from multiple nodes. Simply merging fragmented information makes it difficult to accurately assess the dynamic impact of events on the overall road network, leading to inaccurate priority determination and potentially causing delays in response to critical events or excessive alarms for minor events. This hinders the improvement of the overall effectiveness of the collaborative perception system and the efficiency of command and decision-making. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for real-time push of highway video events based on edge node collaboration.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] This invention discloses a method for real-time push of highway video events based on edge node collaboration, including:

[0008] Acquire local event fragment data generated by each edge node, wherein the local event fragment data includes the spatiotemporal range, motion characteristics and local credibility of the event;

[0009] An event announcement message is generated and broadcast based on the local event fragment data. The event announcement message includes a summary of the motion features and the spatiotemporal range.

[0010] Upon receiving the event announcement message, perform a first-level matching based on the overlap between the spatiotemporal range and the local sensing range, and the summary information; if the matching is successful, obtain detailed local event fragment data from the corresponding node.

[0011] Based on the detailed event fragment data obtained from multiple edge nodes, a second-level matching and time alignment are performed to fuse and generate event chain data and assign an initial fusion confidence level.

[0012] For the time period in the event chain data where the local credibility is lower than the preset credibility threshold, compensation data is obtained from other edge nodes, and the fusion credibility is updated according to the motion continuity between the compensation data and adjacent data.

[0013] The event chain data is mapped to road space grid cells, and the spatiotemporal occupancy intensity of each road space grid cell is calculated and time decay is performed to generate a dynamic space occupancy map.

[0014] Calculate the impact metric of each event chain data based on the dynamic space occupancy map and the fusion credibility.

[0015] The event chain data is sorted according to the impact metric, a push priority instruction is generated, and event notification data is sent.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] 1. This invention generates and broadcasts event announcement messages containing event summary information and spatiotemporal range from each edge node. Upon receiving the message, a first-level matching is performed based on the spatiotemporal overlap and the summary information. Only after a successful match is a detailed event fragment data is obtained and processed. This two-level interaction mechanism avoids the network pressure caused by continuously transmitting raw or high-bandwidth data while achieving collaborative event information perception across multiple nodes. Therefore, with limited edge network resources, it effectively expands the spatiotemporal coverage of collaborative perception and improves the comprehensiveness of event discovery.

[0018] 2. This invention generates event chain data by fusing detailed event fragment data from multiple nodes and assigns it dynamically updated fusion credibility. Simultaneously, it maps the event chain data to a dynamic spatial occupancy graph. This allows the determination of the same event to no longer rely on the local and potentially unreliable observations of a single node, but rather integrates multi-perspective spatiotemporal and motion features for cross-validation and compensation. This improves the accuracy of event detection and the robustness of determining the event's persistence state, and provides a two-dimensional quantitative basis for subsequent evaluation that considers both reliability and spatiotemporal impact.

[0019] 3. This invention calculates a comprehensive impact metric based on fusion credibility and dynamic space occupancy maps, and then sorts and pushes events accordingly. This transforms the decision-making basis for event push from the traditional single event type or local severity to a multi-dimensional comprehensive assessment of the event's reliability, spatiotemporal impact range, and duration. Therefore, it can more accurately identify events that have the most critical impact on road network operational efficiency, allowing limited emergency response resources to be prioritized for the most needed spatiotemporal locations, thus improving the overall efficiency of event response and the level of intelligence in road network management. Attached Figure Description

[0020] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0021] Figure 1 This is a flowchart of the steps of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the working principle of the present invention;

[0023] Figure 3 This is a flowchart of the first-level matching process of the present invention;

[0024] Figure 4 This is a flowchart of the second-level matching process of the present invention. Detailed Implementation

[0025] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0026] Application Overview:

[0027] In existing technologies, real-time processing of highway video events largely relies on centralized cloud platforms or simple front-end detection and reporting models, making it difficult to balance real-time response in edge computing environments with the accuracy of cross-node information coordination. Traditional methods, when faced with limitations in single-camera field of view, occlusion, and differences in clocks and viewing angles between different nodes, result in fragmented and asynchronous reported event information, making it difficult to accurately correlate and fully reconstruct the same event across multiple cameras. Existing systems lack effective assessment of the confidence level of single-node detection, and simple spatiotemporal correlation strategies cannot handle information conflicts caused by viewing angle misalignment and differences in observation conditions, resulting in low reliability of the fused event chain data and making it difficult to support highly reliable command and decision-making.

[0028] To address the aforementioned issues, this study achieves initial matching by exchanging spatiotemporal overlap and feature summary information between nodes. Once a match is successful, detailed event fragment data is acquired for deep fusion, thereby reducing collaborative communication overhead while maintaining real-time performance. Furthermore, it was found that local low-confidence periods observed by a single node can be repaired using compensation data from neighboring nodes. Moreover, utilizing the physical continuity of motion vectors as a verification criterion can effectively identify and fuse high-quality information, dynamically improving the overall confidence of the event chain data.

[0029] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] Example:

[0031] like Figure 1 As shown, the real-time push method for highway video events based on edge node collaboration includes:

[0032] The system acquires local event fragment data generated by each edge node. This local event fragment data includes the spatiotemporal range of the event, motion characteristics, and local credibility. The system generates local event fragment data, including keyframes, bounding box sequences, and displacement and velocity vectors, which are stored in a local cache. Local credibility is calculated in real time by the processor based on image sharpness (gradient magnitude), target occlusion ratio (bounding box overlap rate), and ambient lighting stability parameters, using a scoring algorithm (such as linear weighting).

[0033] Network communication modules (such as 5GCPE, Wi-Fi 6 modules, or Ethernet controllers) on the same edge node read the generated local event fragment data from the cache. Based on the local event fragment data, event announcement messages are generated and broadcast, which include summary information on motion characteristics and spatiotemporal range.

[0034] Other edge nodes in the network (as receiving nodes) continuously monitor the network, receive event announcement messages, and perform first-level matching based on the overlap between the spatiotemporal range and the local sensing range, as well as summary information; if the matching is successful, they obtain detailed local event fragment data from the corresponding node.

[0035] A second-level matching and time alignment is performed based on detailed event fragment data obtained from multiple edge nodes, which are then fused to generate event chain data and assigned an initial fusion credibility.

[0036] Motion continuity checks rely on the calculation and comparison of the average motion vector. The system employs a unified method to calculate the average motion vector: vector fitting (such as linear regression) of the trajectory point sequence to avoid errors caused by simple differencing. The direction change angle is calculated using the vector dot product formula to ensure angle accuracy. Preset parameters such as angle tolerance (15-30 degrees) and speed ratio (25%-40%) are set based on historical traffic flow data statistics to handle lane changes or acceleration / deceleration during normal driving behavior.

[0037] Event chain data is mapped to road spatial grid cells, and the spatiotemporal occupancy intensity of each road spatial grid cell is calculated and time decay is applied to generate a dynamic spatial occupancy map.

[0038] The impact metric for each event chain is calculated based on a dynamic spatial occupancy map and fusion reliability. Specifically, firstly, the total number of road space grid cells or their total area affected by the event chain data in the dynamic spatial occupancy map are counted, quantified as a "spatial coverage parameter." Secondly, the duration of the event chain data from occurrence to termination is obtained, quantified as a "temporal duration parameter." Then, the spatial coverage parameter, temporal duration parameter, and fusion reliability of the event chain data are calculated using a pre-defined weighted summation formula to obtain a comprehensive value, i.e., the impact metric. This value reflects the comprehensive impact of the event on the road network and the reliability of the information.

[0039] The event chain data is sorted according to the impact metric, and the sorting result directly determines the order of push notifications. The system encapsulates the key information (such as event type, location, time, and impact metric) of the first (or top N) event chain data into a structured push priority instruction and sends the event notification data.

[0040] like Figure 2As shown, the working principle of this application is as follows: video streams are continuously acquired through various edge nodes (such as smart cameras equipped with AI chips or roadside computing units). These devices have built-in processors (such as ARM Cortex-A series or NVIDIA Jetson series) and memory (such as DDR4 RAM), running lightweight target detection and tracking algorithms (such as YOLOv4-Tiny or lightweight optical flow algorithms). The processor analyzes the video frames, and when it detects abnormal motion patterns that meet preset conditions (such as vehicle stagnation, driving in the wrong direction, or abnormal speed changes), it triggers event recording.

[0041] The system extracts motion features (e.g., trajectory direction encoding, average speed value) and spatiotemporal range (event start / end timestamps, approximate location range in the road coordinate system), and generates a fixed-length digest using lightweight computation methods such as hash functions (e.g., MD5 or CRC32). Subsequently, the system broadcasts or multicasts the event announcement message containing the digest and spatiotemporal range to pre-configured IP addresses or multicast addresses of neighboring nodes that are physically close or belong to the same logical road segment, using either broadcast or multicast methods (e.g., using UDP).

[0042] Other edge nodes in the network (acting as receiving nodes) continuously monitor the network. When they receive an event announcement message, they parse the spatiotemporal range in the message and calculate its overlap with the local sensing range of their own device (i.e., the fixed field of view of the camera in the road coordinate system). Simultaneously, they compare the received summary information with the event summary information recently generated by themselves or announced by other nodes in their cache, calculating the similarity (e.g., calculating the Hamming distance of hash values). If the calculated spatiotemporal overlap exceeds a preset overlap threshold (e.g., 30%) and the summary similarity exceeds a preset similarity threshold (e.g., 80%), it is considered a successful first-level match. At this point, the receiving node initiates a point-to-point data request to the source node that sent the announcement, requesting detailed local event fragment data corresponding to the event.

[0043] When the source node receives a data request, its communication module retrieves the original detailed event fragment data (typically containing a more complete sequence of trajectory points, original or compressed keyframes, etc.) from its local cache and sends it to the requesting node via the network. By collecting this type of detailed event fragment data from multiple edge nodes, spatiotemporal clustering is performed on them based on the timestamps and calibrated road coordinate system location information carried in each data point, initially grouping different fragments that may describe the same physical event. Subsequently, multiple trajectory time series within the same cluster are precisely aligned on the time axis using a time series alignment algorithm based on dynamic programming (such as Dynamic Time Warping (DTW)). After alignment, the system fuses these temporally calibrated data segments from different perspectives to generate a unified, temporally continuous event chain data. Simultaneously, based on the local credibility of each original data segment participating in the fusion and their average error after spatial alignment, an initial fusion credibility is assigned to the event chain data through weighted calculation or lookup table mapping.

[0044] After generating event chain data, if a time segment originates from an edge node and its local confidence level is lower than a preset, empirical confidence threshold (e.g., below 0.6), it is determined that the information for this time segment may have quality defects (such as severe occlusion). In this case, the system sends compensation data requests via its network interface to other edge nodes within that time segment that may have better viewing angles (determined based on pre-stored node topology and view maps). The edge nodes receiving the requests retrieve observation data for the corresponding time segment from their caches (which may be clearer trajectory points or keyframes) and return it. Upon receiving the compensation data, it first compares it with adjacent data segments before and after that time segment in the event chain data.

[0045] The system maintains a digital grid model corresponding to the physical space of roads (i.e., road space grid cells), with each cell corresponding to a small area in reality (e.g., 10 meters x one lane). The processor maps the event chain data obtained in the previous step (essentially a time-stamped location sequence) onto this grid, marking all road space grid cells traversed by the event. For each occupied road space grid cell, the system calculates an initial spatiotemporal occupancy intensity value based on the duration the event resides in that cell and the number of lanes affected (e.g., occupying two lanes has a greater impact than occupying one lane), and adds it to a time series queue maintained for that road space grid cell. Simultaneously, an independently running timed decay process traverses the queues of all road space grid cells at a fixed period (e.g., per second), performing exponential decay operations on each intensity value based on the time elapsed since the event's occurrence, and removing records that fall below a threshold after decay. The intensity values ​​of all road space grid cells collectively constitute a dynamic spatial occupancy map reflecting the real-time impact of events on the road network.

[0046] Based on the fusion reliability of the dynamic space occupancy map and event chain data, the impact metric for each event chain data is calculated. Finally, the system sorts all event chain data from highest to lowest according to their impact metrics. After generating a push priority instruction, the regional data processing unit sends the instruction and corresponding event summary information (event notification data) via the network to the designated instruction issuing device, such as the control console server of the traffic management center, the variable message sign controller on the road section, or the mobile terminal of the patrol vehicle. Upon receiving the instruction and notification data, these downstream devices can trigger the corresponding display, alarm, or handling procedures.

[0047] Through the specific implementation methods described above for hardware and software collaboration, this application achieves the following: by constructing an occupancy map that dynamically reflects the evolution of event spatial impact and combining it with a comprehensive evaluation of event credibility, the system can automatically and objectively identify events that have the greatest impact on road network operation and require the highest priority for handling, thereby generating differentiated push instructions. This enables traffic managers to make decisions based on a more accurate and comprehensive real-time situational view, thereby improving the intelligence level and overall efficiency of highway event response and road network management.

[0048] like Figure 3 The diagram shows the flowchart for the first-level matching process. This application further proposes that, during the first-level matching process based on the overlap between the spatiotemporal range and the local sensing range, and the summary information, after the collaborative processing unit parses the announcement message from the network packet, it will execute two computational tasks in parallel: calculating the similarity score of the summary information and the percentage of overlapping area of ​​the spatiotemporal range. The system pre-stores the polygon coordinates of the sensing range of the local node's camera in a unified road coordinate system in memory. Simultaneously, the "spatiotemporal range" in the announcement message is typically represented spatially as a rectangular box or a polygonal region (defined by the envelope of the event trajectory). The processor uses a computational geometry library (such as based on the Sutherland-Hodgman algorithm) to calculate the intersection area of ​​these two polygons and divides it by the area of ​​the local node's sensing range to obtain a value between 0 and 1, i.e., the percentage of overlapping area.

[0049] The summary information in the announcement message is typically a fixed-length (e.g., 128-bit) binary string or integer value obtained by hashing the motion features (e.g., using the MurmurHash3 algorithm). The receiving node stores not only its own recently generated event fragments in its local cache, but also summaries of announcement messages received from other nodes, collectively forming the summary information of the motion features. When calculating similarity, the processor compares the received summary information with the summary information of each motion feature in its local cache. A typical implementation is to calculate the Hamming distance between two summary strings, i.e., the number of different bits between them. To obtain a standardized score, it can be converted into a similarity score S, calculated as follows:

[0050]

[0051] in, This is a summary of the received announcement. This is a locally cached summary of motion features. The number of bits in the summary. The similarity score S also ranges from 0 to 1, with a larger value indicating greater similarity in motion patterns between the two events.

[0052] A successful match is determined when the similarity score is greater than a preset first threshold T1 and the overlap area ratio (OR) is greater than a preset second threshold T2. These two thresholds are empirical parameters pre-set based on a large amount of historical road segment data statistics and system performance requirements. The preset first threshold T1 can be set in the range of 0.6 to 0.8 to ensure that the motion pattern has sufficient relevance; the preset second threshold T2 can be set in the range of 0.2 to 0.4 to tolerate a certain amount of projection error and capture events in adjacent or partially overlapping fields of view.

[0053] Through the specific implementation methods described above, this application enables edge nodes to filter out high-probability related neighbor node events with only a minimal amount of metadata (summary and spatiotemporal range). This reduces unnecessary and bandwidth-intensive blind requests for detailed event fragment data, enhancing the adaptability and robustness of the entire system in complex and ever-changing highway environments.

[0054] like Figure 4 The diagram shown is a flowchart of the second-level matching process. This application further proposes to perform second-level matching and time alignment based on detailed event fragment data obtained from multiple edge nodes, and to fuse and generate event chain data and assign an initial fusion credibility. In specific implementation, detailed event fragment data from different edge nodes are clustered based on the timestamps of the detailed event fragment data and their positions in the road coordinate system.

[0055] Each detailed event segment contains a series of timestamped trajectory points and their two-dimensional coordinates (x, y) in a unified road coordinate system (obtained through prior camera calibration). The clustering algorithm is based on the principle of spatiotemporal proximity: the system calculates the combined temporal and spatial distance between any two data segments. A typical distance metric combines time windows and spatial Euclidean distance; for example, if the maximum time difference between two data segments is within Δt (e.g., 5 seconds), and the average spatial distance between their trajectory points is less than D meters (e.g., 50 meters), they are considered potentially related. Based on this metric, density-based clustering algorithms (such as DBSCAN) or hierarchical clustering algorithms can be used to divide all data segments into several clusters.

[0056] For multiple data segments within each cluster (typically from 2-4 different cameras), a dynamic time warping algorithm is used to align the motion trajectories on their time series. Assume a cluster contains M data segments, and each data segment i contains N_i trajectory point sequences. The core of the DTW algorithm is to construct a cumulative distance matrix and find a path from (1,1) to (...). , The optimal curved path is found that minimizes the total distance between corresponding points on the path. Here, the distance between points is usually calculated using spatial Euclidean distance. By performing DTW on pairwise data segments, a one-to-many or many-to-one correspondence between their trajectory points can be established, thereby achieving non-linear stretching or compression on the time axis to achieve alignment.

[0057] An initial fusion confidence score is assigned based on the spatial consistency error of the aligned trajectories; the smaller the error, the higher the initial confidence score. For each aligned time point, multiple location estimates (obtained via DTW path mapping) from different data segments corresponding to that time are fused. A common fusion method is weighted averaging, where the weights can be based on the original local confidence scores of each data segment. Assume there are R aligned location estimates at time k. Its corresponding local trust level is The position after fusion The calculation is as follows:

[0058]

[0059] By concatenating the fused positions at all times, the event chain data for that event is generated. Simultaneously, to evaluate the quality of this fusion and assign an initial fusion confidence level to the event chain data, the system calculates the spatial consistency error. For each time k, the estimated positions of all participating positions relative to the fused position are calculated. The weighted root mean square error. Assume the weights are the normalized local confidence scores. Then the local error at time k for:

[0060]

[0061] Average spatial consistency error of the entire event chain data That is, all valid moments. The average value. This error value intuitively reflects the dispersion of the estimates of the same target location from multiple observation sources. The smaller the dispersion, the higher the consistency of the observations and the more reliable the fusion result. The system is based on... The size of is determined by a pre-defined monotonically decreasing mapping function. To give the initial fusion credibility :

[0062]

[0063] in, This represents the highest confidence level (e.g., 1.0). It is the attenuation coefficient (e.g., 0.1 per meter). The unit is meters. Average spatial consistency error. The typical effective range is between 0.5 meters and 5 meters, depending on the camera's accuracy and distance.

[0064] Through the specific implementation methods described above, this application not only solves the time alignment problem of multi-source heterogeneous observation data, but also generates an objective and interpretable initial credibility index for the fusion results by quantifying spatial consistency errors. This enables the system to synthesize fragmented local observations into a complete event trajectory that is temporally continuous, geographically accurate, and has a reliability metric, providing a high-quality and credible data foundation for subsequent decision analysis. Simultaneously, the error-based credibility assignment method ensures that the system's confidence judgment is based on computable physical consistency, enhancing the scientific rigor and robustness of the entire technical solution.

[0065] This application further proposes that when using the dynamic time warping algorithm to align motion trajectories, its path search strategy is based on the ant colony optimization algorithm, with minimizing the cumulative spatial distance deviation as the optimization objective, thereby determining the optimal alignment path.

[0066] When performing Dynamic Time Warping (DTW) on motion trajectory sequences from different nodes for time alignment, the standard DTW algorithm searches for the optimal solution by dynamically programming through all possible paths. This can incur considerable computational overhead when processing long sequences or requiring numerous pairwise matchings. To improve computational efficiency while maintaining alignment accuracy and adapting to the real-time requirements of edge computing environments, this application proposes using Ant Colony Optimization (ACO) to guide the optimal path search process in DTW. This is an optimization strategy that combines heuristic search with the core ideas of DTW.

[0067] In practical implementation, when processing two trajectory sequences to be aligned within a cluster (let's say sequence A and sequence B, with lengths N and M respectively), the regional data processing unit constructs the DTW path search problem as a graph-based optimal path search problem. This graph is an N×M grid, where each grid node (i,j) represents the probability of matching the i-th point of sequence A with the j-th point of sequence B. Directed edges exist between node (i,j) and subsequent allowed nodes (typically (i+1,j), (i,j+1), (i+1,j+1)), representing the extension of the path. Each edge has two attributes: heuristic information (η) and pheromone concentration (τ). Heuristic information... It is usually defined as the reciprocal of the local matching cost from node u to node v. For example, if u=(i,j) and v=(i+1,j+1), then the local cost is the Euclidean distance between point i+1 in sequence A and point j+1 in sequence B. ,So ε is a very small constant to prevent division by zero. Heuristic information guides ants to tend to choose edges with smaller local distances.

[0068] Ant colony optimization (ACO) simulates the behavior of an ant colony searching for the shortest path in a graph. During algorithm initialization, the pheromone concentration τ on each edge is set to the same initial value. (For example, τ_0 = 1.0). Then, multiple iterations are performed (for example, setting the number of iterations T to be between 50 and 200). In each iteration, a certain number of "ants" (software-simulated agents, typically K between 10 and 50) are released starting from the initial node (1,1). Each ant chooses its next move based on the pheromone concentration on each outgoing edge and heuristic information, according to probability. The probability of moving from node u to node v is... Determined by the following formula:

[0069]

[0070] in, The set of all allowed successors of the current node u is denoted by α. α and β are two key parameters that control the relative influence of pheromone and heuristic information, respectively. α values ​​are typically between 0.5 and 2.0, and β values ​​are typically between 1.0 and 5.0. Larger β values ​​make the search more reliant on local heuristics (i.e., more greedy), while larger α values ​​make the search more inclined to follow paths with higher pheromone concentrations (i.e., stronger convergence). The specific values ​​of these parameters can be calibrated through pre-experiments on typical trajectory datasets to achieve a balance between search efficiency and solution quality.

[0071] After each ant independently constructs its path from (1,1) to (N,M), the total cumulative spatial distance deviation of its path is calculated, which is the sum of the Euclidean distances between any two sequence points corresponding to all nodes traversed along the path. This is precisely the optimization objective that the algorithm aims to minimize. Assume the path of the k-th ant is... The total cost is In this iteration, pheromone updates are performed after all ants have completed their pathfinding.

[0072] After a predetermined number of iterations T, the algorithm terminates. At this point, the path with the highest pheromone concentration (or the best path found historically) is adopted as the final alignment path for DTW. Based on this path, a point-to-point correspondence between sequence A and sequence B can be established, thus completing time alignment.

[0073] By employing the aforementioned ant colony optimization-based path search strategy, compared to standard dynamic programming (DTW), the ACO-DTW hybrid algorithm, through the introduction of heuristic information and the parallel search capabilities of swarm intelligence, can more efficiently approximate the globally optimal alignment path in a larger search space. This is particularly evident when handling long or noisy trajectory sequences, demonstrating better robustness and shorter computation time. This enables faster and more precise time alignment of multiple trajectory sequences even with the limited computing power of edge servers, thus supporting the real-time requirements of the entire event fusion process and improving the system's processing capabilities in high-concurrency event scenarios.

[0074] This application further proposes that the reliability of the fusion update based on the motion continuity of the compensation data and adjacent data includes:

[0075] Calculate the average motion vector of the compensated data within the time period. Calculating the average motion vector is not limited to simple differences between the start and end points. A more robust implementation involves vector fitting to all trajectory points within the time period. For example, for compensated data, a linear regression method can be used, with time as the independent variable and position coordinates as the dependent variable, to fit a line of displacement change within that time period; the slope of this line is the average velocity vector. For event chain data, the same method can be used to calculate vectors within adjacent time periods (e.g., windows ΔT = 3 seconds before and after a missing time period), or the arithmetic mean of all instantaneous velocity vectors within the window can be used. When calculating the angle of change of direction, the vector dot product formula can be used, accurate to the angle value.

[0076] Calculate the average motion vector of the event chain data within adjacent time periods before and after the current time period.

[0077] If the directional change angle of the preceding and following vectors is less than the preset angle tolerance and the rate of change of speed is less than the preset ratio, the motion is considered continuous, and the fusion reliability is improved based on the historical observation accuracy of the edge nodes providing compensation data or the clarity index of the compensation data itself; otherwise, the fusion reliability is reduced. The preset angle tolerance can be set according to the normal driving behavior of vehicles on highways. For example, considering normal lane changes and slight directional adjustments, this tolerance can be set in the range of 15 to 30 degrees. The preset ratio of the rate of change of speed takes into account the possible acceleration and deceleration of vehicles before and after the event, but should not include abnormal situations such as sudden braking or sudden acceleration. This ratio can be set between 0.25 and 0.4 (i.e., allowing a speed change of 25% to 40%). The establishment of these thresholds can be based on statistical analysis of a large amount of normal traffic flow data. For example, calculate the speed and direction change distribution of 95% of vehicles in adjacent short time periods, and take a high quantile value of this distribution (such as 90%) as the upper limit of the tolerance.

[0078] After determining the continuity of motion, there are various calculation models for improving the credibility of fusion based on the historical accuracy of nodes or the clarity of data.

[0079] 1. Weighted improvement model based on historical observation accuracy of nodes: The system maintains a dynamically updated historical observation accuracy score for each edge node. (Initial value can be 0.5). This score is iteratively updated based on the accuracy of the compensation data (or all reported data) provided by the node historically, verified through subsequent processes (such as manual confirmation or comparison with data from other high-precision nodes). When using the compensation data of node n, the confidence improvement is... It can be designed as: ,in This is the gain coefficient (e.g., 0.1). The higher the value is (0.5) above the baseline, the greater the improvement.

[0080] 2. A quantitative improvement model based on the sharpness index of the compensated data itself: Sharpness index This can be calculated directly from the video frames associated with the compensation data. A common method is to calculate the average gradient magnitude (reflecting texture sharpness) and local contrast of the image in the target region. These metrics are then normalized and weighted to obtain the final result. (Between 0 and 1). The credibility enhancement amount can be designed as follows: , where μ is the quality coefficient (e.g., 0.15).

[0081] 3. Integrated Enhancement Model: A more comprehensive approach is to combine the two methods mentioned above, for example: ,in The weights can be adjusted based on experience or online learning. The final updated fusion credibility... Ensure that the limit is not exceeded.

[0082] When motion continuity verification fails (direction or velocity changes exceed tolerance), the fusion confidence level needs to be reduced. A simple approach is to apply a fixed penalty value. (e.g., 0.1). More refined methods can be correlated with the degree of discontinuity; for example, assuming the direction exceeds the tolerance by an angle of... The speed exceeds the ratio. Then the penalty amount ,in , As the penalty coefficient, the system presets two thresholds for judgment:

[0083] Angle tolerance threshold This threshold defines the permissible range of abrupt changes in direction of motion. Considering normal lane changes or slight steering on highways, this threshold is typically set between 10 and 30 degrees.

[0084] Speed ​​ratio threshold This threshold defines the permissible range of speed variation. Considering the continuity of acceleration and deceleration, this threshold is typically set between 0.2 and 0.5 (i.e., 20% to 50%).

[0085] After the update To avoid drastic fluctuations in reliability due to single-moment misjudgments or data noise, a smoothing mechanism can be introduced. For example, before the final update, the calculated... or Perform low-pass filtering, or require consistent verification results (e.g., 2 times) before performing a significant confidence adjustment.

[0086] Through the above technical solutions, this application enables the validity verification of compensation data to shift from qualitative to quantitative methods, improving the objectivity and consistency of the judgment. The multi-model credibility enhancement strategy based on node historical performance and data quality allows the system to more precisely evaluate the value of compensation data from different sources, thus reflecting the reliability of the fused information more fairly and accurately. The introduced smoothing or anti-jitter mechanism enhances the system's stability in the presence of data noise.

[0087] This application further proposes that, in the specific implementation of calculating the spatiotemporal occupancy intensity of each road spatial grid unit and performing time decay, the regional data processing unit maintains a two-dimensional grid data structure in memory that precisely corresponds to the physical road digital map. Each "road spatial grid unit" corresponds to a fixed area in the real world (e.g., 10 meters in length × the width of a single lane). The system maintains an intensity value queue for each road spatial grid unit. Newly generated intensity values ​​are added to the intensity value queue after being weighted based on the duration of the event chain data and the number of lanes affected. The calculation formula is:

[0088]

[0089] in, This indicates a new contribution value that will soon be added to the target road spatial grid cell intensity value queue;

[0090] This represents the current fusion confidence level of the event chain data. It is a value between 0 and 1, derived from the output of the previous stage. Introducing this factor means that high-confidence events have a higher confidence weight on their impact on the road network;

[0091] This represents the local duration (in seconds) of the event chain data within the current road space grid cell; the system accurately calculates the dwell time of the vehicle within this specific road space grid cell based on the event chain data trajectory data.

[0092] This represents the number of lanes affected globally by the event chain data. For example, an accident occupying two lanes has a significantly larger impact than a temporary stop occupying only one lane. L is typically a positive integer (e.g., 1, 2, 3). The system determines this value by analyzing the set of lanes to which the road space grid cells covered by the event chain data trajectory belong.

[0093] Calculated newly generated strength value The current timestamp will be appended. Then, it is added to the tail of the intensity value queue of the corresponding road space grid cell.

[0094] To simulate the process of influence diminishing over time, the system periodically (e.g., every second) or each time a new effect is observed. After joining, an exponential decay operation is performed on the intensity value queue of all road space grid cells. For each record in the queue... its decayed value Based on the current system time Calculate using the following formula:

[0095]

[0096] in, It is the attenuation coefficient, a preset parameter greater than 0, which determines the rate attenuation. The unit is Its value needs to be adjusted according to the actual application requirements, and the typical value range is within... arrive Between these two values, there are slower (approximately 11.5 minutes half-life) and faster (approximately 1.15 minutes half-life) decay rates, respectively.

[0097] It is the difference between the current time and the time when the record of this intensity was generated (in seconds).

[0098] After the decay process is completed, the system iterates through the queue and permanently removes records whose decayed strength value is lower than a certain set clear threshold (e.g., 0.01) from the queue to save storage space and keep the queue clean.

[0099] Finally, in order to obtain the real-time spatiotemporal occupancy intensity of any road space grid cell at the current moment... The system only needs to process all the attenuated intensity values ​​in the unit queue. Overlay:

[0100]

[0101] Where n is the number of records in the current queue. The value represents the comprehensive impact of historical and current events accumulated in that road spatial grid cell, and decaying over time. All road spatial grid cells... The values ​​together constitute a dynamic spatial occupancy map that reflects the real-time status snapshot of the entire road network.

[0102] This application transforms discrete event trajectory data into a continuous, dynamic, and quantifiable road network "impact heatmap" through physical grid-based quantization accumulation and a time decay model conforming to natural laws. This map not only reflects the location where the event is occurring but also reveals the diffusion and dissipation process of the event's ongoing impact in time and space. This allows the system to move beyond a simple reliance on instantaneous event types and instead assess the actual impact of events on traffic flow from a more macroscopic and dynamic perspective. It provides an objective, continuous, and computable data foundation for subsequent accurate calculation of event impact metrics and prioritization, thereby enhancing the depth of traffic situational awareness and the intelligence level of decision support.

[0103] This application further proposes that, when calculating the impact metric of each event chain data, the calculation is based on the dynamic spatial occupancy graph and the fusion credibility, specifically as follows:

[0104] The total number or area of ​​road space grid cells affected by the current event chain data in the dynamic space occupancy map is quantified as a spatial coverage parameter. The quantification of the spatial coverage parameter is not limited to simple road space grid cell counts or total area calculations. A more refined implementation can consider the spatial distribution of influence intensity. For example, the sum or average of the spatiotemporal occupancy intensity values ​​of all road space grid cells affected by the event chain data in the dynamic space occupancy map can be calculated, reflecting the "concentration" of the influence rather than just its "range." Another approach is to define the spatial coverage parameter as the number of road space grid cells whose occupancy intensity exceeds a certain threshold (such as a high-intensity threshold), focusing on the core congestion area caused by the event. For the "duration parameter," in addition to directly using the total duration of the event (in seconds), a non-linear mapping of the duration can be considered, such as taking the logarithm of the duration or using a piecewise function, to reflect the diminishing marginal effect of duration on the road network (e.g., the first 5 minutes of stagnation have a significant impact, but the marginal impact slows down with each subsequent minute).

[0105] Quantize the duration of the current event chain data into a time duration parameter;

[0106] The impact metric is a weighted sum of spatial coverage parameters, temporal duration parameters, and fusion credibility, with its weight coefficients dynamically adjusted by a particle swarm optimization algorithm. The particle swarm optimization algorithm uses the actual traffic management effect feedback after historical event push as the fitness function to iteratively solve for the optimal combination of weight coefficients.

[0107] The basic formula for calculating the metric I is expressed as a weighted sum of the spatial coverage parameter (S), the temporal duration parameter (T), and the fusion confidence (C), i.e. ,in , , These are weighting coefficients, and they usually satisfy the normalization condition ( Weight initialization can be based on domain experience; for example, the initial settings might focus more on spatial influence and credibility (e.g., ...). =0.4, =0.2, =0.4).

[0108] Finding the optimal combination of weight coefficients , , Defined as an optimization problem in a continuous space. Each particle represents a possible weight vector. The fitness function is the objective function of the PSO algorithm, used to evaluate the quality of a set of weights. In practice, the system records the impact metrics (calculated based on the weights at that time) of all pushed events (especially high-priority events) over a past period (e.g., a week), the content of the push instructions, and key performance feedback indicators. These feedback indicators can be automatically obtained through integration with the traffic control system. For a given weight combination, the impact metrics of historical events are recalculated using these metrics, and the historical events are sorted according to these values.

[0109] Assess the correlation or consistency between this new ranking and the "ideal ranking" based on actual effect feedback. One feasible method for calculating the fitness F is to examine whether events with poor actual mitigation effects (such as long dissipation times) receive higher impact metrics (i.e., are ranked higher) in the new ranking. This ranking consistency can be quantified using metrics such as the Spearman rank correlation coefficient, and this correlation coefficient can be used as the fitness value F. The optimization objective is to maximize F.

[0110] The particle swarm optimization (PSO) algorithm initializes a certain number (e.g., 20-50) of particles within the feasible solution space of the weight coefficients (each weight is between 0 and 1, and their sum is 1). Each particle has a random position (weight vector) and velocity. In each iteration:

[0111] a. For each particle, calculate its fitness value using the weight vector at its current position, as described above.

[0112] b. Each particle records its own historical best position.

[0113] c. The entire particle swarm records the global historical best position found among all particles.

[0114] d. Update the velocity and position of each particle according to the particle swarm optimization formula. The velocity update formula considers the particle's own experience, social experience, and inertia. The position update involves moving within the solution space guided by the velocity.

[0115] e. Constrain the updated positions to ensure that the weight coefficients satisfy normalization and non-negativity.

[0116] The algorithm terminates when the preset maximum number of iterations (e.g., 100) is reached, or when the global optimal fitness no longer significantly improves over multiple generations. The resulting global optimal position is then considered a new set of optimal weight coefficients. The system updates the weights in the online impact metric calculation module to this new set of values. This optimization process can be performed offline periodically (e.g., weekly), enabling the model to dynamically self-adjust.

[0117] Through the aforementioned technologies, the system can automatically learn how to configure the relative importance of spatial coverage, temporal duration, and credibility to the true urgency of an event under different traffic conditions and road segment characteristics. The resulting impact metric and the corresponding push priority can better align with the dynamic needs of actual road network management, thereby continuously improving the accuracy of event push notifications and the overall effectiveness of traffic emergency response.

[0118] This application further proposes that, after sorting the event chain data according to the impact metric and before generating the push priority instruction, a first-level verification is also included, the specific steps of which include:

[0119] The system sets a verification window size N (e.g., N=5 or 10) to examine the top N event chains, extracting the local spatial influence range corresponding to each of these N events. It analyzes whether there are continuous, high-intensity road spatial grid cells forming a congestion zone in the corresponding dynamic spatial occupancy map. In this application, a congestion zone is defined as: within the event influence imprint range, there exists a strip-shaped or sheet-like region of a certain length and width, composed of continuous (i.e., spatially adjacent) and high-intensity road spatial grid cells. The criterion for determining "continuous" is to cluster adjacent high-intensity road spatial grid cells using a four-connected or eight-connected image region growing algorithm. The criterion for determining "high intensity" is the current spatiotemporal occupancy intensity of the road spatial grid cells. Exceeding a preset high intensity threshold High intensity threshold The setting should be based on the intensity of influences that typically lead to a significant decrease in vehicle speed or the formation of queues in historical data, and the value should be within the top 20% quantile of the overall intensity value distribution. For example, through statistical analysis, a high intensity threshold can be established. An empirical value between 1.5 and 3.0 can be set. Additionally, to eliminate scattered high-intensity points, the system requires that the total number X of continuous high-intensity road space grid cells within the identified "blockage zone" area be greater than a minimum size threshold. (For example, And its length-to-width ratio or extension length meets certain conditions (for example, the length exceeds 50 meters).

[0120] Based on the verification results, the system dynamically adjusts the priority of events:

[0121] If this exists, it indicates that the event not only has a significant impact on itself, but its impact has also created a coherent and large-scale congestion zone in space, which is highly likely to have already triggered or is about to trigger a chain reaction of congestion in upstream vehicles, resulting in global and transmissive damage to the road network's capacity. Therefore, the system will further increase the priority of this event chain data. The magnitude of the increase can be positively correlated with parameters such as the size M of the congestion zone and its average intensity. For example, the new priority score... The original impact measurement value can be used. Add a reward item to the basics: ,in It is a reward coefficient (e.g., 0.1).

[0122] If the current event chain data exhibits a sporadic distribution in its dynamic spatial occupancy map—meaning that while events may have high impact metrics (potentially due to long duration or affecting multiple lanes), their high-intensity cells in the dynamic spatial occupancy map are discrete and discontinuous, failing to form significant clusters—then the priority is downgraded according to a preset priority rule based on the dispersion of the current event chain data. The dispersion level D can be measured using the standard deviation of the spatial coordinates of the high-intensity road spatial grid cells, or the average nearest neighbor distance. A larger D indicates a more dispersed distribution. The downgrade rule could be: a new priority score... ,in It is a suppression coefficient (e.g., 0.05). The preset priority rule ensures that the greater the dispersion, the greater the downsampling.

[0123] Through the above specific implementation methods, the final push instructions generated by this application are more in line with the practical needs of overall road network traffic management and efficiency maximization, avoiding the decision-making system being misled by events with high numerical values ​​but low actual harm. Thus, in complex concurrent highway event scenarios, more accurate and intelligent emergency resource scheduling and information dissemination guidance are achieved.

[0124] This application further proposes that a differentiated push strategy be adopted for sending event notification data. In specific implementation, the system will classify event chain data according to a preset priority grading standard (e.g., dividing the scores affecting metric ranking or after verification and adjustment into high, medium, and low levels). For events of different levels, the following differentiated push process will be executed:

[0125] 1. For the highest priority events: These are typically events that pose a serious or urgent threat to the operation of the road network (such as serious traffic accidents, large-scale congestion, fires, etc.), requiring multi-party coordination and rapid response. Therefore, the system adopts a multi-party synchronous and comprehensive information push method.

[0126] The announcement content includes a notification containing complete data. This notification not only includes a basic text summary of the event (such as event type, precise location, time, and affected lanes), but also embeds keyframe images (i.e., one or more frames extracted from the event video stream that best reflect the event situation, usually compressed) and trajectory prediction information (such as the possible location of vehicles or the spread trend of their impact range in the next few seconds based on their current motion state). This combination of information provides traffic managers with an intuitive view of the scene, provides clear evidence for information board dissemination and vehicle warnings, and provides on-site personnel with predictive information.

[0127] 2. For medium-priority events: These are events with potential impact or requiring monitoring but not yet reaching an emergency level (e.g., slow traffic, partial road closures, minor debris spills, etc.). These typically require attention and record-keeping by central personnel. Notification content: Announcements will be sent in the form of a text summary. The content will be concise, containing only essential event attribute information such as type, location, status, and initial impact metric. This avoids consuming large amounts of data such as images and predictions from non-urgent events, thus limiting the limited bandwidth and storage resources of the central and front-end systems and ensuring smooth communication channels for high-priority events.

[0128] 3. For low-priority events: those with minor impact and likely to dissipate quickly (such as a bicycle briefly stopping and then leaving, or a minor traffic violation that has ended), there is usually no need to immediately trigger a central alarm or front-end announcement. No proactive push to the central system or any front-end is made. The system only retains the notification data for such events (a concise text summary and related metadata) in the local storage medium (cache) of the edge node or regional data processing unit that generated the event for a period of time (e.g., 1 hour).

[0129] When traffic management personnel actively query historical event records for specific time periods and road sections through the management platform, the system then retrieves and provides summary information of these low-priority events from the cache.

[0130] Through the above technical solutions, this application enables the entire system to maintain its focus on core threats and efficient information flow in the complex and ever-changing environment of highways with a large number of concurrent events, thereby improving the overall operational efficiency and robustness of the traffic incident emergency management system.

[0131] This application further proposes that, after generating push priority instructions and sending event notification data, the method also includes an offline learning step, specifically including:

[0132] The system collects all processed historical event chain data within a specific time period (e.g., the past month) from the long-term archive database of the online system. It also collects the final fusion reliability of the historical event chain data and the labels indicating whether the historical event chain data has been subsequently manually verified.

[0133] For each historical event chain, its complete multi-dimensional features are collected. These features typically include:

[0134] Spatiotemporal characteristics: such as the total duration of the event, the number or area of ​​the overall spatial road grid units affected, the highest spatial occupancy intensity, and the average movement speed.

[0135] Collaborative observation characteristics include: the number of original edge nodes involved in the fusion, the average historical observation accuracy of these nodes, and the average spatial consistency error calculated during data fusion.

[0136] Credibility feature: This refers to the final fusion credibility value of the event chain data in the online processing flow.

[0137] Manual Confirmation Tag: This is a crucial monitoring signal. The system obtains a Boolean tag from the event handling logs of the traffic control center via an interface, indicating whether each historical event chain has been subsequently manually confirmed (e.g., verified on-site by traffic police or confirmed as a genuine event through multi-channel video review).

[0138] Using the dataset collected above ,in It is the multi-dimensional feature vector of the i-th event chain data. It is the credibility of its final fusion. The labels are manually verified (1 indicates true, 0 indicates false alarm or cannot be verified), and the system trains a classification model. The model's input is the feature vector. With final fusion credibility Enhanced feature vectors formed by combination The output is the predicted probability that the event is a real event. .

[0139] The model can choose algorithms suitable for structured data and with strong interpretability, such as gradient boosting decision trees or logistic regression. The training dataset contains 10,000 historical event chains, with 20-dimensional features extracted from each event chain, including spatiotemporal features, co-observation features, and confidence features. LightGBM is used to implement gradient boosting decision trees, with 200 trees and a maximum depth of 6, employing early stopping to prevent overfitting. The training process iteratively builds multiple decision trees, with each tree learning the residuals of the predictions from all previous trees. The number of trees is typically between 100 and 500, with early stopping to prevent overfitting. The maximum tree depth is limited to 3 to 8 to control model complexity and enhance generalization ability. The learning rate is set to 0.01, the subsampling ratio is 0.8, and logarithmic or exponential loss is used. Training selects optimal parameters through cross-validation, aiming to minimize the predicted probability. With real labels The gap between them.

[0140] After training, the optimal classification model M is obtained. This model is encapsulated and deployed back into the regional data processing unit of the online system as a confidence calibration module. Subsequently, whenever the online system generates a new event chain data, its initial fusion confidence is calculated. Or the fusion credibility after compensation and update At that time, the system will simultaneously perform calibration operations:

[0141] Extract the same multi-dimensional feature vectors from the new event chain data as during training. The feature vector is then compared with the current confidence value to be calibrated. Combine to form input The data is fed into the classification model M. It outputs a predicted probability. This probability representation model, based on historical patterns, determines the likelihood that the event corresponding to the current event chain data is a real event.

[0142] System utilization right Perform calibration and use the calibration function to obtain the calibrated confidence value. The calculation formula is as follows:

[0143]

[0144] in, It is a smoothing coefficient (e.g., 0.3) used to control the calibration amplitude and avoid drastic fluctuations. When If the model deems the event more credible based on historical patterns, the credibility score is increased; otherwise, it is decreased. (Calibrated) It will replace the original credibility value and be used for subsequent impact metric calculations and priority ranking.

[0145] Through the above technical solution, this application can effectively correct the credibility assessment bias caused by specific scenarios, rare event types, or factors not fully covered by online rules. This makes the system's output credibility not only based on real-time physical consistency verification and node quality assessment, but also incorporate statistical patterns extracted from a large number of historical success and failure cases. This improves the overall accuracy of event authenticity judgment and the system's adaptability to complex and ambiguous scenarios, making the final push decision based on a more solid and intelligent confidence foundation.

[0146] The following is a specific embodiment of a method for real-time push of highway video events based on edge node collaboration:

[0147] A collaborative perception network consisting of five edge nodes (smart cameras) was deployed on a section of a highway. Each node has a coverage radius of approximately 500 meters, and the sampling frequency is 25 frames per second. During the morning rush hour one day, edge node 1 (K1) detected a white SUV abnormally stagnating in the emergency lane for more than 15 seconds, generating event segment data containing the spatiotemporal range (08:23-08:25, coordinates X=12345, Y=6789), motion characteristics (trajectory stagnation, speed <5km / h), and local confidence level (0.72). Node 1 broadcast an event announcement message via UDP protocol, containing a spatiotemporal range summary (08:23-08:25, X±50m, Y±30m) and motion characteristic summary information (trajectory stagnation code H01). After receiving the announcement, neighboring node 2 (K2) calculates the spatiotemporal overlap (40% overlap between the perception ranges of node 1 and node 2) and the summary similarity (85% matching degree of H01 encoding), determines that the first-level match is successful, and then sends a detailed event fragment data request to node 1.

[0148] Node 1 sends detailed event fragment data, including a complete trajectory point sequence (20 timestamped coordinates), a keyframe image (frame 08:23:15), and local confidence level, to Node 2. Node 2, combined with its synchronously detected slow-moving events in adjacent lanes (08:22-08:26, speed <20km / h), performs a second-level matching and time alignment on the two sets of data: a dynamic time warping algorithm is used to align the stationary trajectory (20 seconds) of Node 1 with the slow-moving trajectory (25 seconds) of Node 2, calculating a time offset Δt = 1.25 seconds and a spatial consistency error of 0.8 meters. Cross-node event chain data is then generated through fusion, with an initial fusion confidence level. Calculated based on the number of participating nodes (2) and the consistency error: (Where 0.8 is the DTW alignment error percentage, and 50 is the preset maximum error threshold).

[0149] After mapping the event chain data to the road spatial grid (10 meters × lane width), an anomaly was found to cover 3 road spatial grid cells. The spatiotemporal occupancy intensity was calculated as follows: =0.6×(3 grids / 5 total grids)+0.4×0.8=0.68 (0.6 is the spatial coverage weight, 0.4 is the time decay coefficient). Based on the particle swarm optimization algorithm (PSO), the weight coefficients are dynamically adjusted, and the current event impact metric is: P=0.76×0.68+0.24×(1-0.15)=0.62 (0.76 is the fusion confidence, 0.24 is the spatial occupancy weight, and 0.15 is the time decay threshold). After comparison with the historical event database, it is determined to be a medium priority event (above the threshold of 0.5 but below the high priority threshold of 0.7).

[0150] After receiving the event chain data synchronously, node 3 (K3) initiates congestion zone analysis: It extracts road space grid cells (5 consecutive high-intensity cells, occupancy intensity > 0.8) from the abnormal area, revealing that the SUV's stoppage has caused the spatiotemporal occupancy intensity of three lanes within a 200-meter radius behind it to rise to 0.75 (safe threshold 0.6). A comprehensive impact metric is calculated based on the data from the three nodes. (0.6 is the three-node coordination weight, 0.3 is the congestion intensity weight, and 0.1 is the time decay coefficient). The system generates push priority instructions: send a complete announcement containing the key frame (08:23:15 frame) and trajectory prediction to the traffic control center; push a text warning "There is a vehicle malfunction in the emergency lane ahead, please drive carefully" to the road information board; and push dispatch information containing location coordinates and handling suggestions to the patrol car terminal 1 kilometer away from the incident point.

[0151] Subsequent offline learning trains a classification model based on historical data. Inputting multi-dimensional features of the event chain data (duration of stagnation, scope of impact, number of participating nodes), the system outputs a predicted probability of 0.89 for the actual occurrence of this type of event. When similar event announcements with a spatiotemporal range (±10%) and motion characteristics (encoding matching degree >80%) are detected again, the system will automatically trigger a cross-node compensation data request, further improving the credibility of the event chain data fusion to above 0.85.

[0152] This embodiment demonstrates how multiple edge nodes can solve the problem of limited vision of a single node through lightweight collaboration. It integrates the originally fragmented local events into a complete chain of road anomaly events. Through dynamic weighted impact measurement calculation and priority adjustment of particle swarm optimization, it ensures that event push focuses on the core congestion area while avoiding excessive interference from non-critical events, ultimately improving the accuracy and efficiency of highway event handling.

[0153] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for real-time push of highway video events based on edge node collaboration, characterized in that: include: Acquire local event fragment data generated by each edge node, wherein the local event fragment data includes the spatiotemporal range, motion characteristics and local credibility of the event; An event announcement message is generated and broadcast based on the local event fragment data. The event announcement message includes a summary of the motion features and the spatiotemporal range. Upon receiving the event announcement message, perform a first-level matching based on the overlap between the spatiotemporal range and the local sensing range, and the summary information; If a match is found, detailed event fragment data is retrieved from the corresponding node; Based on the detailed event fragment data obtained from multiple edge nodes, a second-level matching and time alignment are performed to fuse and generate event chain data and assign an initial fusion confidence level. For the time period in the event chain data where the local credibility is lower than the preset credibility threshold, compensation data is obtained from other edge nodes, and the fusion credibility is updated according to the motion continuity between the compensation data and adjacent data. The event chain data is mapped to road space grid cells, and the spatiotemporal occupancy intensity of each road space grid cell is calculated and time decay is performed to generate a dynamic space occupancy map. Calculate the impact metric of each event chain data based on the dynamic space occupancy map and the fusion credibility. The event chain data is sorted according to the impact metric, a push priority instruction is generated, and event notification data is sent.

2. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: The specific steps for performing the first-level matching based on the overlap between the spatiotemporal range and the local sensing range and the summary information include: Calculate the percentage of overlap between the similarity score of the summary information and the spatiotemporal range; When the similarity score is greater than a preset first threshold and the overlap area ratio is greater than a preset second threshold, the match is considered successful. The similarity score is obtained by comparing the summary information with the locally cached motion feature summary information.

3. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: The specific steps for performing second-level matching and time alignment based on the detailed event fragment data obtained from multiple edge nodes, fusing to generate event chain data, and assigning initial fusion confidence include: For the detailed event fragment data from different edge nodes, correlation clustering is performed based on the timestamps of the detailed event fragment data and their positions in the road coordinate system; For each data segment within a cluster, a dynamic time warping algorithm is used to align its motion trajectory on the time series. The initial fusion confidence level is assigned based on the spatial consistency error of the aligned trajectory.

4. The method for real-time push of highway video events based on edge node collaboration according to claim 3, characterized in that: When aligning motion trajectories using the dynamic time warping algorithm, the path search strategy is based on the ant colony optimization algorithm, with minimizing the cumulative spatial distance deviation as the optimization objective, thereby determining the optimal alignment path.

5. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: The reliability of the motion continuity update fusion based on the compensation data and adjacent data includes: Calculate the average motion vector of the compensated data within the time period; Calculate the average motion vector of the event chain data within adjacent time periods before and after the stated time period; If the directional change angle of the preceding and following vectors is less than a preset angle tolerance and the rate of change of velocity is less than a preset ratio, then the motion is determined to be continuous, and the fusion reliability is improved based on the historical observation accuracy of the edge nodes providing the compensation data or the clarity index of the compensation data itself; otherwise, the fusion reliability is reduced.

6. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: Calculating the spatiotemporal occupancy intensity of each road spatial grid cell and performing time decay includes: An intensity value queue is maintained for each road space grid cell. Newly generated intensity values ​​are added to the intensity value queue after being weighted based on the duration of the event chain data and the number of lanes affected. For each intensity value in the intensity value queue, it is exponentially decayed according to the timestamp of its corresponding event; The attenuated intensity values ​​are superimposed to form the current spatiotemporal occupancy intensity of the road space grid cell.

7. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: When calculating the impact metric for each event chain, the calculation is based on the dynamic space occupancy map and the fusion confidence level, specifically as follows: The total number or total area of ​​road space grid cells affected by the current event chain data in the dynamic space occupancy graph is quantified into a space coverage parameter. Quantize the duration of the current event chain data into a time duration parameter; The impact metric is a weighted sum of the spatial coverage parameter, the temporal duration parameter, and the fusion credibility, and its weight coefficients are dynamically adjusted by a particle swarm optimization algorithm. The particle swarm optimization algorithm uses the actual traffic management effect feedback after the historical event push as the fitness function to iteratively solve for the optimal combination of weight coefficients.

8. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: After sorting the event chain data according to the impact metric and before generating the push priority instruction, a first-level verification is also included, the specific steps of which include: Examine the top N event chain data and analyze whether there are continuous, high-intensity road space grid units forming a blockage zone in the corresponding dynamic space occupancy map; If it exists, the priority of the current event chain data will be further increased; if the dynamic space occupancy graph of the current event chain data is sporadically distributed, the priority will be reduced according to the dispersion of the current event chain data and the preset priority rules.

9. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: The sending of event notification data adopts a differentiated push strategy, specifically including: For the highest priority events, a complete notification containing keyframes and trajectory predictions is simultaneously pushed to the traffic control center, road information boards, and mobile terminals upstream of the affected lanes. For medium-priority events, only a text summary is sent to the traffic control center; For low-priority events, they are cached only on the edge side for on-demand querying.

10. The method for real-time push of highway video events based on edge node collaboration according to claim 1, characterized in that: After generating push priority instructions and sending event notification data, the method further includes an offline learning step, specifically including: Collect historical event chain data and the final fusion credibility of the corresponding historical event chain data, as well as the label of whether the historical event chain data has been subsequently manually confirmed; Based on the collected historical event chain data and corresponding labels, a classification model is trained, and based on the multi-dimensional features of the event chain data, the classification model is used to obtain the predicted probability of the actual occurrence of the event corresponding to the event chain data. The predicted probability is used as prior knowledge to calibrate the initial or updated fusion confidence.

Images