Road video event rapid detection method based on edge computing

Abstract

The application discloses a road video event fast detection method based on edge calculation, and particularly relates to the field of computer vision and image processing, and is used for solving the motion evidence distortion problem caused by frame-level time sequence disorder and content gap in road monitoring video. The method is achieved by reading a video segment, constructing a time sequence correction graph according to interframe content similarity and boundary continuity, rearranging a corrected sequence by generating a frame consistency index, estimating a stable motion field and a continuity score on the corrected sequence, and writing back a stable image amplitude correction time anchor point. According to the stable motion field, an event evidence sequence is generated, a track breakpoint is merged according to the time sequence correction graph, backtracking revision is initiated, a candidate event segment is returned, consistency verification is performed on the candidate event, if there is a conflict, a time sequence correction graph and a stable image amplitude are revised across levels, a stable motion field is re-estimated, an output confirmed event is triggered, a trigger result is published, a mapping index is established as a new segment initialization priori, and acceleration time sequence correction and judgment convergence are achieved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing, and more specifically, to a method for rapid detection of road video events based on edge computing. Background Technology

[0002] Road surveillance relies on pole-mounted cameras and edge-side devices working in tandem. During video acquisition, compression, transmission, and presentation, issues arise such as frame rearrangement, slight timestamp drift, and brief congestion. Progressive exposure and micro-vibrations of the high-position carrier introduce subtle deformations, while privacy occlusion and dynamic areas of interest alter the visible content and texture details. Edge-side devices rely on continuous motion cues within short time sequences for rapid judgment. However, once the frame order and content continuity shift, early signs are scattered across adjacent segments, resulting in inconsistent motion trajectories and geometric orientations, thus undermining the triggering criteria for short-term judgments.

[0003] However, the core obstacle of existing technologies lies in the distortion of motion evidence caused by frame-level temporal discrepancies and content gaps: the arrival order of the video is inconsistent with the actual occurrence order, the continuity of pixel-level motion is interrupted, and key clues such as trajectory and speed cannot form a stable correlation within a short time sequence, thus losing reliable basis for rapid judgment. It is necessary to introduce content-oriented consistency correction and evidence rearrangement at the image analysis stage to restore the observation sequence to a form close to the actual occurrence order, realign short-term motion features with event triggering logic, and restore the temporal stability and judgment usability upon which edge-side rapid detection relies.

[0004] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a fast road video event detection method based on edge computing. This method reads road video segments, constructs a temporal correction map based on inter-frame content similarity and boundary continuity, rearranges the segments to generate a corrected sequence, and outputs a frame consistency index. It estimates a stable motion field and continuity score on the corrected sequence and writes back the image amplitude correction time anchor point. Based on the stable motion field, it generates an event evidence sequence and initiates backtracking revisions by merging trajectory breakpoints according to the temporal correction map, returning candidate event segments. It performs consistency verification on the candidate events; if there is a conflict, it revises the temporal correction map and image amplitude across levels, triggering a stable motion field re-estimation and outputting a confirmation event. It publishes the triggering result and establishes a mapping index as a priori initialization for new segments, accelerating temporal correction and decision convergence, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] S1: Read road video clips, construct a temporal correction map based on the similarity of inter-frame content and the continuity of boundaries, rearrange the arrival order according to the correction map to generate a corrected sequence, and output the frame consistency index;

[0008] S2: Estimate the stable motion field and continuity fraction on the corrected sequence, write the stable image amplitude back to the time-series correction map to correct the time anchor point of the occlusion interval, and provide a time reference for subsequent evidence generation;

[0009] S3: Generate an event evidence sequence based on the stable sports field, merge trajectory breakpoints according to the time-series correction diagram, and simultaneously initiate a backtracking revision request for the continuity score and return the updated candidate event fragments;

[0010] S4: Perform consistency verification on candidate events. If there is a conflict with the frame consistency index, revise the timing correction map and image stabilization amplitude across levels and trigger the stabilization motion field reassessment. After completing the reassessment, output the confirmation event.

[0011] S5: Publish the trigger result and establish a mapping index between the confirmed event, the event evidence sequence, and the time-series correction graph. This index serves as the initialization prior and cross-level recall basis for new segments, and is used to accelerate the time-series correction and decision convergence of subsequent segments.

[0012] Furthermore, step S1 takes a video segment as input, calculates the overlap of texture histograms and edge bands between adjacent frames to obtain two types of observable consistency signals, constructs a temporal correction graph with each frame as a node, and weighs edges based on a joint sorting of content similarity and boundary continuity, directly eliminating weak consistency edges. A restricted shortest path search is performed on the graph, with path constraints satisfying both timestamp adjacency and structural continuity, resulting in the most coherent frame sequence as the corrected sequence. A frame consistency index is generated based on the corrected sequence, and the index records the preceding and following adjacencies of frames, available intervals, and conflict markers, forming a dense retrieval entry point for temporal verification.

[0013] Furthermore, step S2 performs layered block matching on the corrected sequence, estimates the global micro-vibration trajectory using the distant still area, subtracts the trajectory from the displacement field to obtain the stable motion field, fits the linear drift of the background displacement along the line direction to address the inter-line time difference caused by line-by-line exposure, completes line direction compensation, suppresses rolling shutter deformation, and generates a continuity score using two consistency measures: the directional consistency rate of the same point within the short window and the scarcity of curvature change along the trajectory. The two are fused in order to form a single score, quantifying temporal coherence. The stable image amplitude calculated from the background block is written back to the temporal correction map, and the time anchor point of the occluded interval is updated so that the graph structure and motion estimation share the same time reference.

[0014] Furthermore, step S3 extracts three types of atomic evidence based on the stable motion field, pointing to velocity mutation, path offset, and spatial occupancy change, respectively. The three types of evidence are described by occurrence time, persistence marker, and local geometric relationship. The adjacency and time anchor information of the time-series correction map are called to perform shape and texture matching in a small range before and after occlusion. The trajectory breakpoint caused by occlusion is bounded to be continued, generating an event evidence sequence. For segments with low continuity scores, a backtracking revision request is triggered. The request content includes frame interval, displacement mode, and conflict type, indicating that the motion and time anchor points in the original segment are re-estimated. The evidence after continuation is summarized, and candidate event segments are output, along with the corresponding continuity scores and evidence list.

[0015] Furthermore, step S4 performs consistency verification on candidate event segments. The frame consistency index is used to check the consistency relationship between the frame order and the main path on the graph. The continuity score is used to check whether the motion direction and geometric change are in the same direction. If a conflict is detected, local relaxation is performed on the time-series correction graph. The local frame order is rearranged in a limited manner according to the joint sorting of content similarity and boundary continuity. The amplitude of the image stabilization is fine-tuned based on the displacement residual of the background block. After the revision is completed, the stable motion field is re-estimated and the continuity score is checked only in the affected interval. The event evidence sequence is aligned with minimal changes and re-evaluated to form a confirmed event.

[0016] Furthermore, step S5 generates and publishes the trigger result for the confirmed event, and establishes a mapping index. The index uses the event identifier as the key and maps it to the set of frames participating in the judgment, the set of effective edges of the temporal correction graph, the list of atomic evidence of the event evidence sequence, and the continuity score interval. When a new segment enters the process, scene similarity retrieval is performed according to the mapping index. If the retrieval hits, the corresponding time anchor point and stable image amplitude are directly retrieved as the initial value. Frame order correction and motion estimation are preferentially carried out in similar graph paths and spatial bands to reduce reconstruction costs and shorten convergence time.

[0017] Furthermore, in step S1, the texture distribution vector of each frame is extracted between adjacent frames using the local binary pattern histogram method, the texture histogram similarity is calculated using the cosine similarity technique, the edge pixel set of each frame is extracted using Canny edge detection, and the edge band overlap is calculated using the Hausdorff distance technique to obtain two types of observable consistency signals.

[0018] Furthermore, in step S1, a temporal correction graph is constructed with each frame as a node. For any pair of nodes, the edge weight is determined by the joint sorting of content similarity and boundary continuity. The similarity and overlap are normalized and sorted to obtain the rank value. The two rank values ​​are added together to obtain the joint sorting value, and its reciprocal is taken as the edge weight. If the joint sorting value exceeds the preset threshold, it is determined to be a weakly consistent edge and is removed.

[0019] Furthermore, step S2 performs hierarchical block matching on the corrected sequence, dividing each frame into image blocks of fixed size, identifying distant still areas, using the Lucas-Kanade optical flow method to calculate inter-block displacement vectors for blocks within the distant still areas, obtaining global micro-vibration trajectories, calculating the displacement field of the entire frame, and subtracting the global micro-vibration trajectories from the initial displacement field block by block to obtain a stable motion field.

[0020] Furthermore, in step S2, the background region is identified as having a displacement amplitude in the stable motion field that is lower than the average displacement of the sequence. The background displacement is extracted along the row direction of each frame. The least squares method is used to fit a linear drift model for the row-level displacement sequence. The fitted linear drift is applied to each row of pixels for compensation. The stable motion field is updated so that the displacement of all rows is consistent with the compensated value, thus suppressing the rolling shutter deformation.

[0021] The technical effects and advantages of the fast road video event detection method based on edge computing of this invention are as follows:

[0022] This invention effectively restores the temporal coherence of road video segments by constructing a temporal correction map, rearranging frame order, and generating a frame consistency index, thereby reducing motion cue fragmentation caused by transmission congestion and timestamp drift. It estimates stable motion fields and continuity scores and writes back stabilized image amplitude correction time anchors to suppress camera micro-vibrations and rolling shutter deformation, ensuring the accuracy of motion estimation and shared time reference. It generates event evidence sequences and initiates backtracking revisions by merging trajectory breakpoints according to the temporal correction map, bridging occluded trajectories and improving the integrity and reliability of candidate event segments. It performs consistency verification and, in case of conflicts, revises the temporal correction map and stabilized image amplitude across levels to trigger local re-estimation, forming confirmation events and enhancing the robustness and accuracy of the judgment. It publishes the trigger results and establishes a mapping index as a priori for new segment initialization, accelerating the temporal correction and judgment convergence of subsequent segments and reducing reconstruction costs. This achieves a content-oriented temporal correction and evidence rearrangement closed loop, significantly improving the real-time performance and resource efficiency of rapid road video event detection in edge computing environments, and solving the core obstacle of motion evidence distortion in existing technologies. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the fast detection method for road video events based on edge computing according to the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0025] Example 1: Figure 1 This invention presents a fast road video event detection method based on edge computing, comprising:

[0026] S1 reads road video clips, constructs a temporal correction map based on the similarity of inter-frame content and the continuity of boundaries, rearranges the arrival order according to the correction map to generate a corrected sequence, and outputs the frame consistency index.

[0027] S2 estimates the stable motion field and continuity fraction on the corrected sequence, writes the stable image amplitude back to the time-corrected map to correct the time anchor point of the occlusion interval, and provides a time reference for subsequent evidence generation.

[0028] S3 generates an event evidence sequence based on the stable sports field, merges trajectory breakpoints according to the time-series correction diagram, and simultaneously initiates a backtracking revision request for the continuity score and returns the updated candidate event fragments.

[0029] S4 performs consistency verification on candidate events. If there is a conflict with the frame consistency index, it revises the timing correction map and image stabilization amplitude across levels and triggers a re-evaluation of the stable motion field. After completing the re-evaluation, it outputs a confirmation event.

[0030] S5 publishes the trigger results and establishes a mapping index between the confirmed event, the event evidence sequence, and the time-series correction graph. This index serves as the initialization prior and cross-level recall basis for new segments, accelerating the time-series correction and decision convergence of subsequent segments.

[0031] In road surveillance environments, video acquisition, compression, transmission, and presentation often lead to frame order disorder, timestamp drift, and interruption of content continuity. These problems, coupled with privacy occlusion and dynamic changes in the area of ​​interest, further distort motion cues, making it impossible to reliably determine events in short-time sequences at the edge.

[0032] When road video clips reach the edge, frame-level temporal discrepancies and content gaps distort motion evidence, causing trajectory and velocity cues to be unstablely correlated within a short time series. By reading video clips and constructing a temporal correction map based on inter-frame content similarity and boundary continuity, the frame order can be rearranged to generate a corrected sequence, and a frame consistency index can be output to restore temporal stability.

[0033] The specific processing logic of step S1:

[0034] Using video clips as input, the texture histogram and edge band overlap between adjacent frames are first calculated to obtain two types of observable consistency signals. The video clips, as the initial input, contain a sequence of frames that may have temporal discrepancies, and the content consistency of adjacent frames needs to be quantified to prepare for the construction of the temporal correction map.

[0035] For any pair of adjacent frames f in a video segment i and fi+1 First, texture histogram similarity is calculated, and then the texture distribution vector for each frame is extracted using the local binary pattern histogram method. and in Indicates frame f i Binary pattern frequency vectors of the mid-pixel neighborhood; then, texture histogram similarity s. t Calculated using the cosine similarity formula: Where the dot product · represents the vector inner product, and the norm ∥·∥ is the Euclidean norm. Next, the edge overlap is calculated, and Canny edge detection is used to extract the edge pixel set e for each frame. i and e i+1 Then, the overlap ratio is calculated using the Hausdorff distance formula. Where d(A, B) represents the maximum and minimum distances from set A to set B, and w and h are the frame width and height, respectively, evaluating the geometric alignment of edge structures. Finally, the texture histogram similarity s t overlap with edge band s e As two types of observable consistency signals, we obtain texture and edge consistency quantization signals for adjacent frames in road video clips, providing basic data for edge trade-offs in temporal correction maps and ensuring that the signals directly correspond to inter-frame content similarity and boundary continuity.

[0036] A temporal correction graph is constructed with each frame as a node. The edge weighting is based on a joint ranking of content similarity and boundary continuity, with weakly consistent edges being directly removed. The obtained texture histogram similarity s is then used as the basis for the correction. t overlap with edge band s e Each frame of the video clip needs to be treated as a node to construct a timing correction graph in order to capture the potential correct sequence.

[0037] f of each frame of the video clip k (where k = 1 to the total number of frames n) are used as nodes in the timing correction graph; for any pair of nodes (f p f q The weight w of the edge p，q Based on the joint ranking calculation of content similarity and boundary continuity, firstly, for s t and s e Perform normalized sorting to obtain their respective rank values ​​r. t and r e (Rank values ​​from 1 to the number of possible edges, arranged in descending order of consistency), then the joint sort value j p，q =r t +r e The weight w of the edge p，q =1 / j p，q (Lower rank values ​​correspond to higher weights, promoting coherent paths); if the joint ranking value j p，qIf the edge exceeds a preset threshold (e.g., the 80th percentile of the total possible edges), it is identified as a weakly consistent edge and is directly removed to eliminate low-reliability connections.

[0038] The temporal correction graph is constructed, and its edge weights uniformly reflect the joint ordering of content similarity and boundary continuity. After removing weakly consistent edges, the graph structure focuses on highly reliable frame associations, providing a simplified temporal model for path search. A constrained shortest path search is performed on the graph, with path constraints requiring both timestamp adjacency and structural continuity to be satisfied, resulting in the most coherent frame sequence as the corrected sequence. The constructed temporal correction graph and its edge weights w are then used... p，q It is necessary to search for a coherent path covering all nodes to rearrange the frame order and generate the corrected sequence.

[0039] A constrained shortest path search is performed using a variant of Dijkstra's algorithm, starting from the first timestamp node, with the goal of finding the shortest total weight path covering all nodes. Path constraints include timestamp adjacency (e.g., the timestamp difference between adjacent nodes does not exceed 1.5 times the average frame interval of the video segment) and structural continuity (e.g., the edge weight w between adjacent nodes). p，q The weight of each edge must not exceed 1.2 times the average edge weight in the graph. Both conditions must be met simultaneously to ensure the path is valid. During the search process, edges that satisfy the dual constraints are expanded first to obtain the complete path with the smallest cumulative weight. The node sequence corresponding to this path is the most coherent frame sequence, which is used as the corrected sequence.

[0040] A corrected sequence is generated through restricted path search. The frame order of this sequence has been optimized based on the temporal correction map, restoring the temporal coherence of road video segments and providing a reliable foundation for motion estimation. A frame consistency index is generated based on the corrected sequence. This index records the adjacent frames, available intervals, and conflict markers, forming a dense retrieval entry point for temporal verification. A frame consistency index needs to be created based on the corrected sequence to record the optimized frame relationships, facilitating temporal verification.

[0041] For each frame f in the corrected sequence m Record its preceding and following adjacent frames (previous frame ID and next frame ID), and the available interval (e.g., the length of a continuous high-consistency frame segment starting from this frame, using s). t and s e The average value of the segment exceeds 0.7) and the conflict marker (if the frame does not match the timestamp in the original video segment, it is marked as 1, otherwise it is marked as 0); the index is organized in the form of a hash table, with the frame ID as the key and the above triplet as the value, forming a dense retrieval entry point, supporting O(1) time-series verification query.

[0042] Once the frame consistency index is generated, its records uniformly correspond to the frame relationships of the corrected sequence, ensuring an efficient temporal reference entry point in road video event detection.

[0043] Step S1 restores the frame sequence coherence of the video clip by constructing a temporal correction map and generating a corrected sequence and frame consistency index. Stable motion fields and continuity scores still need to be estimated on the corrected sequence to further suppress the effects of micro-vibrations and deformations, and the stabilized image amplitude is written back to the temporal correction map, thus providing an accurate time reference for event evidence generation.

[0044] The specific processing logic of step S2:

[0045] Layered block matching is performed on the corrected sequence. First, the global micro-vibration trajectory is estimated using the distant still area, and then the trajectory is subtracted from the displacement field to obtain the stable motion field. Based on the corrected sequence generated in step S1, inter-frame displacement needs to be analyzed through layered block matching to isolate global micro-vibrations and obtain a clean motion representation, targeting the subtle jitter caused by high-position carriers in road monitoring.

[0046] First, each frame of the corrected sequence is divided into fixed-size image blocks, for example, each block is 32x32 pixels. Then, a distant still area is identified, which is the upper third of the frame. This area usually contains the sky or distant static objects, and its motion is assumed to be mainly caused by camera micro-vibrations. For the blocks in the distant still area, the Lucas-Kanade optical flow method is used to calculate the inter-block displacement vector. By iteratively solving the velocity equation under the assumption of constant brightness, the global micro-vibration trajectory is obtained, which is the average vector sequence of the displacements of all distant blocks in the sequence. Next, the displacement field of the entire frame of the corrected sequence is calculated. The Lucas-Kanade optical flow is also used to obtain the displacement vector of each block to form the initial displacement field. The global micro-vibration trajectory is subtracted block by block from the initial displacement field, that is, the displacement vector of each block is subtracted from the average vector at the corresponding time, thereby eliminating the influence of camera shake and obtaining a stable motion field. This motion field only reflects the real motion of objects, which is convenient for line compensation and event extraction.

[0047] The Lucas-Kanade optical flow method is a classic computer vision algorithm used to estimate the motion velocity and direction of pixels in an image sequence. Its core is based on the assumption of constant brightness, meaning that the brightness of pixels remains unchanged over a short period. The displacement vector is calculated by solving the optical flow constraint equations. This method applies least-squares optimization to pixels within a local neighborhood, assuming that adjacent pixels have similar motion patterns. This results in high computational efficiency and robustness when dealing with sparse or dense optical flow, making it particularly suitable for real-time applications in video tracking, motion analysis, and other fields.

[0048] A stable motion field is generated through layered block matching and global micro-vibration subtraction. This removes interference from high-level carrier micro-vibrations in road video clips, ensuring that the motion representation focuses on the actual object dynamics. This provides an accurate displacement reference for suppressing rolling shutter deformation and facilitates the calculation of continuity fractions. To address the inter-line time difference caused by progressive exposure, a linear drift of the background displacement is fitted along the line direction to achieve line-direction compensation and suppress rolling shutter deformation. Using the stable motion field, compensation needs to be made for the common inter-line time difference in progressive exposure cameras to correct vertical deformation and adapt to the distortion caused by rapidly moving objects in road monitoring.

[0049] First, background regions are identified, specifically pixel rows in the stable motion field whose displacement amplitude is less than, for example, 10% of the sequence average displacement. These rows typically correspond to static road surfaces or background scenery. Then, background displacement is extracted along the row direction of each frame, i.e., the horizontal and vertical components of background pixels are selected from the stable motion field to form a row-level displacement sequence. For the row-level displacement sequence, a linear drift model is fitted using the least squares method, assuming that the displacement changes linearly with the row number. The fitting parameters include slope and intercept, and the best-fit line is obtained by solving the normal equation. For each row of pixels, the fitted linear drift is compensated by subtracting the fitted value from the row's displacement vector, thereby correcting the offset caused by the inter-row time difference. Finally, the stable motion field is updated to ensure that the displacement of all rows is consistent with the compensated value, suppressing rolling shutter distortion and ensuring the geometric continuity of frame content in the vertical direction.

[0050] After the orientation compensation is completed, the stable motion field has eliminated the linear drift deformation caused by line-by-line exposure, which improves the accuracy of pixel layer motion in the processing of road video clips, provides a distortion-free motion basis for generating continuity scores, and supports the reliable extraction of evidence sequences.

[0051] A continuity score is generated using two consistency measures: the directional consistency rate of the same point within a short window and the scarcity of curvature change along the trajectory. These two measures are fused in an ordered manner to form a single score used to quantify temporal coherence. Based on the compensated stable motion field, the temporal coherence of motion needs to be quantified. A continuity score is generated by fusing the two measures to evaluate the stability of motion cues in the corrected sequence, addressing discontinuities caused by privacy occlusion.

[0052] First, calculate the orientation consistency rate. This involves calculating the similarity of the displacement directions of each pixel within a short time window (e.g., 5 frames) in a stable motion field, using the cosine of the vector angles to accumulate the similarity. For each pixel p within the window, the sequence of displacement vectors... Directional consistency rate in Let represent the displacement vector of the i-th frame. A normalization factor ensures the score is between 0 and 1, capturing local directional stability through directional consistency rather than simple averaging. Secondly, the curvature change scarcity is calculated, i.e., the frequency of curvature change is evaluated along the pixel trajectory (formed by accumulated displacement from a stable motion field), using a finite difference approximation of curvature: for a sequence of trajectory points... curvature Then scarcity Where θ is the curvature threshold set based on the average curvature of the sequence, and scarcity is used to quantify the smoothness of the trajectory; finally, a single continuous score is generated through sequential fusion, namely the direction consistency rate c. d and curvature variation rarefaction c r After sorting, the geometric mean c is taken to ensure that the fusion emphasizes the complementarity of the two and is used for overall temporal coherence assessment.

[0053] Summary after processing: After the continuity score is generated, it integrates direction and curvature measures, accurately quantifies the motion coherence of the corrected sequence in the road monitoring scenario, provides a threshold basis for retrospective revision, and enhances the reliability of time anchor point correction of the time series correction map.

[0054] Step S2 estimates the stable motion field and continuity score on the corrected sequence and writes the stable image amplitude back to the time-series correction plot to unify the time reference. To generate a complete event evidence sequence, atomic evidence needs to be extracted based on the stable motion field, trajectory breakpoints need to be merged, and backtracking revisions need to be initiated for low continuity segments, thereby outputting updated candidate event fragments to support subsequent verification.

[0055] The specific processing logic of step S3:

[0056] Three types of atomic evidence are extracted from the stable motion field, pointing to abrupt changes in velocity, path deviation, and changes in spatial occupancy, respectively. These three types of evidence are described by their occurrence time, persistent markers, and local geometric relationships. Based on the stable motion field generated in step S2, key motion anomalies need to be extracted as atomic evidence to capture core clues of potential events in road videos, such as sudden vehicle stops or changes in occupancy, addressing dynamic inconsistencies within short time series.

[0057] First, a pixel-level scan of the stable motion field is performed to identify evidence of abrupt velocity changes, i.e., to detect sharp changes in the magnitude of the displacement vector: for the displacement sequence of each pixel trajectory. Calculate acceleration approximation as in Let a represent the displacement vector of the k-th frame, where Δt is the frame interval time. kIf the acceleration exceeds twice the threshold of the sequence average, it is marked as evidence of a velocity abrupt change, and the occurrence time (corresponding frame timestamp), duration marker (number of consecutive frames exceeding the threshold), and local geometric relationship (average displacement direction of the surrounding 8 neighboring pixels) are recorded. Next, path offset evidence is extracted and evaluated through trajectory deviation: a fitted straight line for the pixel trajectory is constructed, and the position sequence is fitted using least squares. Then calculate the perpendicular distance from each point to the fitted line. in To fit the line direction vector, if the maximum distance exceeds a threshold such as 10% of the trajectory length, it is marked as evidence of path offset, along with the occurrence time, duration, and local geometric relationship (angular deviation of the offset direction). Finally, evidence of spatial occupancy change is extracted based on regional density variation: the stable motion field is divided into grid regions, and the proportion of moving pixels in each grid is calculated. Where I(·) is the indicator function, ∈ is the small displacement threshold, and area l For each grid area, if the difference in area between adjacent frames exceeds, for example, a 20% threshold, it is marked as evidence of spatial occupancy change, and the occurrence time, continuous marking, and local geometric relationship (a brief description of the boundary contour of the changing grid) are recorded. These three types of atomic evidence are uniformly stored in the form of structured tuples for easy merging later.

[0058] After the atomic evidence extraction is completed, the three types of evidence have been accurately separated and described from the stable motion field. In road video event detection, abnormal patterns of speed, path, and occupancy are highlighted, providing fine-grained clues for merging trajectory breakpoints. Simultaneously, the temporal and geometric attributes of the evidence description support efficient continuation operations, reducing the risk of evidence loss due to occlusion. The adjacency and time anchor information of the temporal correction map are invoked to perform shape and texture matching within a small range before and after occlusion, performing bounded continuation of the trajectory breakpoints caused by occlusion to generate an event evidence sequence. Using the extracted three types of atomic evidence and the temporal correction map written back in step S2 (including updated time anchors), the breakpoint trajectories need to be merged to bridge occlusion interruptions, addressing trajectory fragmentation caused by dynamic occlusion in road monitoring.

[0059] First, the adjacency information (frame IDs before and after the frame) and time anchors (corrected time references) of the temporal correction map are queried to identify occlusion intervals, i.e., frame segments with continuity scores below the average. Then, within a small window of, for example, 3 frames before and after the occlusion, shape matching is performed on the trajectory endpoints in the atomic evidence. Hu invariant moments are used to describe the local contours of the endpoints, and the moment vector similarity m between the endpoints before and after the occlusion is calculated. s =1-∑|log(hu) pre,i / hu post，i )| / 7, where hu pre，i and hu post，iFor the i-th Hu moment, scale rotation is avoided; simultaneously, texture matching is performed, local texture features are extracted using the gray-level co-occurrence matrix, and the Euclidean distance between the contrast and correlation differences is calculated as the matching score; if both the shape and texture matching scores exceed, for example, a threshold of 0.8, bounded continuation is performed, that is, the discontinuity displacement vector is linearly interpolated under the constraint of time anchor point, and the interpolation length is limited to not exceeding, for example, 1.5 times the number of occluded frames; finally, the continuation atomic evidence is sorted according to the time of occurrence and compiled into an event evidence sequence, which contains a continuous chain of all three types of evidence.

[0060] After trajectory breakpoint reconnection is completed, the event evidence sequence integrates clues before and after occlusion, restoring the integrity of the motion trajectory in road video segment processing. This provides a continuous evidence basis for revising low-continuity segments. Simultaneously, anchor points in the temporal correction map ensure the temporal accuracy of the reconnection, enhancing the robustness and efficiency of event candidate generation. For segments with low continuity scores, a backtracking revision request is triggered. The request includes the frame interval, displacement mode, and conflict type, indicating a re-estimation of motion and temporal anchor points in the original segment. Based on the event evidence sequence and the continuity score from step S2, revisions must be initiated for segments with scores below a threshold to correct potential motion estimation errors and adapt to local inconsistencies caused by congestion or drift in road video.

[0061] First, the frames corresponding to the event evidence sequence are scanned to identify segments with low continuity scores, i.e., segments with scores lower than the overall median of the sequence for three or more consecutive frames. Then, a backtracking revision request is generated, which includes, in a structured format, the frame interval (start and end frame IDs), displacement pattern (a summary of the average displacement vector direction and amplitude of the stable motion field within the segment), and conflict type (classified as velocity inconsistency, path curvature, or occupancy anomaly, determined based on the dominant type of atomic evidence). The request instructs a return to step S2 to re-evaluate motion, i.e., re-perform hierarchical block matching and line direction compensation within the specified frame interval, using the displacement pattern as an initialization constraint. Simultaneously, it instructs an update of the time anchor point, i.e., adjusting the anchor point value of the time series correction map according to the revised stable image amplitude. The revision request is sent through a queue mechanism to ensure that it is only executed locally for low-score segments, avoiding recalculation of the entire sequence.

[0062] After the retrospective revision request is triggered, potential errors in low-continuity segments are identified and reassessed. This improves the reliability of the evidence sequence in the road monitoring event detection process, provides corrected motion and timing data for summarizing candidate event segments, reduces false positives due to initial estimation bias, and optimizes the overall accuracy and response speed. The subsequent evidence is summarized, and candidate event segments are output, along with corresponding continuity scores and evidence lists. Integrating the event evidence sequence, revision results, and continuity scores, evidence needs to be summarized to form candidate segments for initial screening of multi-clue events in road videos.

[0063] First, the revised evidence sequence is grouped into three categories of atomic evidence based on occurrence time and local geometric relationship. The DBSCAN algorithm is used for clustering based on spatiotemporal distance (Euclidean distance plus time difference), and evidence within a cluster is considered as the same event segment. Then, the boundary of each cluster is extracted as a candidate event segment, with the starting frame being the earliest occurrence time and the ending frame being the latest continuity marker. A corresponding continuity score, i.e., the average score of frames within the cluster, and an evidence list, i.e., listing the type, time, and description of all atomic evidence in the cluster, are also provided. Finally, the candidate event segment list is output, sorted by the start time, for easy verification in step S4.

[0064] Once the candidate event fragments are output, they have been summarized with continuation evidence and accompanied by scores and lists. This provides a structured preliminary event representation in fast road video detection, laying a comprehensive foundation for consistency verification. At the same time, revision and integration ensure the temporal coherence and evidence integrity of the fragments, significantly reducing the computational burden and probability of misjudgment in subsequent processing.

[0065] Step S3 generates an event evidence sequence based on the stable motion field and outputs candidate event segments through merging breakpoints and backtracking revisions. To ensure the accuracy of the segments, consistency verification is also performed on the candidate events, and in case of conflict, the timing correction map and image stabilization amplitude are revised across levels to trigger local reassessment. Finally, confirmed events are output to achieve accurate and rapid detection of road videos.

[0066] The specific processing logic of step S4:

[0067] S4-1. Perform consistency verification on candidate event segments. First, use the frame consistency index to check the consistency between the frame order and the main path on the graph. Then, use the continuity score to check whether the motion direction and geometric changes are in the same direction. Based on the candidate event segments output in step S3 and their accompanying continuity scores and evidence list, the frame consistency index from step S1 needs to be used for verification to check the temporal and motion reliability of the segments in the road video, and to address any residual errors.

[0068] Extract the frame sequence table of candidate event segments, i.e., the frame ID sequence within the segment ordered by start time; then, use the frame consistency index to query the preceding and following adjacencies and available intervals of each frame, comparing whether the segment frame order matches the main path of the time-series correction map (i.e., the sequence obtained from the restricted shortest path search in step S1), specifically by checking whether adjacent frame IDs are consistent and whether the available interval covers the segment length; if they do not match, they are marked as order conflicts; secondly, for the stable motion field data within the segment, calculate the motion direction consistency, i.e., the displacement vector sequence of atomic evidence in the evidence list. Find the average cosine of each pair of included angles in Let be the displacement vector of the i-th piece of evidence. This formula ensures the overall directional stability of the quantification. At the same time, the homogeneity of geometric changes is evaluated by comparing the local geometric relationship (contour or offset angle) of the atomic evidence with the continuity score. If the score is higher than the segment average and the deviation of the geometric change angle is less than, for example, a threshold of 30 degrees, it is considered to be homogeneous. Otherwise, it is marked as a motion conflict. If any order conflict or motion conflict exists, the verification is deemed to have failed and the process is moved to revision.

[0069] After the consistency verification is completed, the temporal and motion relationships of candidate event segments have been accurately checked through frame consistency index and continuity score. Potential conflict areas are identified in road video event detection, providing targeted triggering basis for subsequent cross-level revisions. At the same time, it ensures that only conflict segments enter the reprocessing path, optimizes the allocation of edge computing resources, improves the overall credibility of the judgment, and avoids the risk of false event propagation.

[0070] S4-2. If a conflict is detected, local relaxation is performed on the time-series correction map. The local frame order is rearranged in a limited manner based on the joint sorting of content similarity and boundary continuity. The amplitude of the stabilized image is fine-tuned based on the displacement residual of the background block. When a conflict is detected in step S4-1, the time-series correction map (the stabilized image amplitude has been written back in step S2) needs to be returned to for local adjustment. The joint sorting of content similarity and boundary continuity in step S1 is used to correct the frame order error for the conflicting segment. The amplitude is fine-tuned in combination with the background block data of the stabilized motion field to bridge the local inconsistencies in road monitoring.

[0071] The process involves locating the temporal correction subgraph corresponding to the conflicting segments, extracting the nodes and edges of the involved frames, performing local relaxation (relaxing edge weight constraints within ±20% of the original value for increased flexibility), re-evaluating edges based on a joint sorting of content similarity (texture histogram similarity) and boundary continuity (edge ​​overlap), calculating and updating the joint sorting value (rank summation) for subgraph edges, selecting the top 50% of high-ranking edges for finite rearrangement (limiting the rearrangement depth to no more than twice the number of fragment frames), and generating new local paths by replacing edges one by one using a greedy algorithm. Simultaneously, the displacement residuals of the background blocks are calculated, representing the difference sequence between the actual displacement of background pixels in the stable motion field and the fitted linear drift. Finally, the amplitude of the image stabilization is fine-tuned by adding a residual correction δ to the amplitude field of each conflicting frame. a =∑r b / n b , where r b For a single background block residual, n b The number of blocks is calculated by accumulating the average residuals to finely adjust the amplitude, ensuring alignment with motion estimation; the local structure and amplitude values ​​of the time series correction plot are updated after revision.

[0072] After local relaxation and fine-tuning, the conflict intervals of the temporal correction map have been optimized through joint sorting finite rearrangement and residual amplitude adjustment. In the road video processing scenario, the local coherence of the frame order and the accuracy of the time anchor points have been restored, providing a refined graph foundation for stable motion field reestimation. At the same time, the computational overhead of full map reconstruction has been reduced, improving the adaptability and efficiency of the detection method, and providing an effective remedy for local distortions caused by transmission congestion or occlusion.

[0073] S4-3. After revision, stable motion field reassessment and continuity score verification are performed only in the affected area. Based on the revised temporal correction map and image stabilization amplitude, the stable motion field needs to be reassessed within the affected area of ​​the conflict segment (e.g., 2 frames before and after expansion). The continuity score is verified using the hierarchical block matching method in step S2 to verify the revision effect and correct the remaining motion deviation in the road video.

[0074] First, the affected interval is defined as the conflict segment frames plus the preceding and following buffer frames. Then, the stable motion field is re-estimated, that is, the global micro-vibration trajectory estimation and displacement field subtraction of the distant quiet zone are re-performed within the interval, and the revised image stabilization amplitude is used as the initialization vector. The Lucas-Kanade optical flow is applied to calculate and update the displacement vector. Next, the orientation compensation is performed, and the linear drift along the background displacement is fitted and subtracted. The continuity score is checked, that is, the orientation consistency rate and curvature change scarcity are recalculated for the updated stable motion field, and then sequentially fused into a single score. If the new score improves more than the original value, for example, by 15%, the revision is confirmed to be effective. Otherwise, the local relaxation of step S4-2 is iterated until the target is met. The whole process is limited to the affected interval to avoid global recalculation.

[0075] After reassessment and verification, the stable motion field and continuity score of the affected section have been accurately updated based on the revised time-series correction map, which enhances the stability of motion representation in the rapid detection of road monitoring events, provides reliable verification data for the alignment of event evidence sequences, significantly reduces computational complexity through local constraints, ensures the real-time performance of the method on edge devices, and provides a closed-loop correction mechanism for micro-vibration or drift residues, thereby improving the overall event confirmation accuracy.

[0076] S4-4. The event evidence sequence is then aligned with minimal modifications and re-evaluated; consistency results in a confirmed event. Using the re-estimated stable motion field and continuity score, the event evidence sequence from step S3 needs to be aligned and adjusted to re-evaluate candidate event segments with minimal modifications. For evidence mismatches in the road video, a final confirmed event is formed.

[0077] First, compare the atomic evidence lists of the event evidence sequences before and after the revision to identify discrepancies (items with inconsistent timing or geometric changes); perform minimal modification alignment, i.e., only adjust the occurrence time and persistence markers of the discrepancies to match the displacement patterns of the re-estimated stable motion field, using a dynamic time warping algorithm to align the sequences, limiting the number of modifications to, for example, no more than 10% of the total number of evidence; then, re-evaluate the segments, i.e., based on the aligned sequences and the verification continuity score, check whether the conflict-free condition is met (frame order matches the main path and motion is in the same direction); if consistent, mark the segment as a confirmed event, and update the evidence list and score accordingly; otherwise, loop back to sub-step S4-2 until convergence; output the list of confirmed events, sorted by event identifier.

[0078] After alignment and re-evaluation, the event evidence sequence has been incorporated into the revision results with minimal changes, forming a confirmed event. This ensures the final consistency and integrity of the evidence in the road video detection process, provides high-quality output for index building in step S5, and improves the robustness and decision convergence speed of the method by handling stubborn conflicts through an iterative mechanism. It achieves accurate event confirmation for complex scenarios such as multi-object interaction or continuous occlusion, and optimizes the response time of edge computing.

[0079] Step S4 ensures reliable confirmation of event segments by outputting confirmation events through consistency verification and cross-level revision. However, to achieve a closed-loop detection process and improve efficiency, it is also necessary to publish trigger results and establish a mapping index between confirmation events and intermediate quantities as initial priors for new video segments, thereby accelerating temporal correction and decision convergence.

[0080] The specific processing logic of step S5:

[0081] Generate and publish trigger results for confirmed events. Based on the confirmed event and its accompanying event evidence sequence and continuity score output in step S4, a structured trigger result needs to be generated to notify edge devices or upper-level modules of road events, addressing the immediate response needs in monitoring scenarios such as vehicle anomalies.

[0082] The system compiles a list of atomic evidence confirming the event, including descriptions of velocity mutations, path offsets, and spatial occupancy changes. Then, it generates a trigger result, which encapsulates the event identifier (a unique string, such as a hash based on timestamps and frame IDs), event type (categorized by dominant evidence type), severity (a continuity score inversely mapped to a 1-10 scale, with higher scores indicating lower severity), and location coordinates (the geometric center extracted from a stable sports field) in JSON format. The trigger result is then published and sent to the monitoring center or associated device via a message queue protocol, ensuring a real-time latency of no more than 100 milliseconds. Simultaneously, a publication log is recorded, including the sending time and confirmation receipt, to support auditing.

[0083] Once the trigger result is generated and published, the event is confirmed to have been transformed into an actionable alarm signal. In the road video surveillance environment, the event is reported in real time, providing accurate type and location information for operational response. At the same time, it bridges the output of the detection process, ensuring that subsequent segment processing can inherit the event context, reducing the burden of repetitive calculations and improving the overall system's alert efficiency.

[0084] Simultaneously, a mapping index is established, with the event identifier as the key, mapping to the set of frames involved in the judgment, the set of effective edges of the temporal correction graph, the list of atomic evidence of the event evidence sequence, and the continuity score interval. Utilizing the confirmed event and its components (such as the event evidence sequence in step S3 and the revised temporal correction graph in step S4), a persistent index needs to be built to store multiple intermediate quantity associations, avoiding reconstruction from scratch to meet the reuse requirements of continuous segments of road video.

[0085] A hash table index is created using the event identifier as the primary key; the mapped values ​​include the frame set (a list of all frame IDs involved in the confirmed event, arranged in order), the effective edge set of the time-series correction graph (edge ​​ID and weight pairs on the revised path, retaining only the strong edges before joint sorting, for example, 30%), the atomic evidence list of the event evidence sequence (a complete description array of the three types of evidence, sorted by the time of occurrence), and the continuity score interval (the minimum and maximum score pairs within the segment); the index is stored in the key-value database and supports O(1) queries; to ensure integrity, a checksum field is added, which is calculated as a concatenated hash of all mapped values.

[0086] Once the mapping index is established, it confirms that events and key intermediate quantities have formed an efficient association structure, which facilitates the rapid retrieval of historical data in road video event detection, provides event-level prior knowledge for the initialization of new segments, and ensures the accurate inheritance of timing and motion during reuse by preserving effective edges and score intervals, significantly reducing the resource consumption of edge computing and strengthening the continuity of cross-segment judgment.

[0087] When a new segment enters the process, scene similarity retrieval is performed according to the mapping index. If the retrieval is successful, the corresponding time anchor point and image stabilization amplitude are directly retrieved as initial values. Frame order correction and motion estimation are then prioritized within similar graph paths and spatial bands to reduce reconstruction costs and shorten convergence time. When a new road video segment is input, retrieval is performed based on the mapping index, utilizing scene feature matching of confirmed events to accelerate the construction of the temporal correction graph in step S1 and motion estimation in step S2, targeting similar scenes in continuous monitoring streams, such as repeated intersection events.

[0088] Scene features of the new fragment are extracted, including the average texture histogram and edge band overlap as vector representations; similarity retrieval is performed, i.e., cosine similarity is calculated for all event identifiers of the mapped index. in For the feature vector of the new segment, The event feature vector in the index is used. If the similarity exceeds a threshold of, for example, 0.75, it is considered a hit. If it is a hit, the corresponding time anchor point (the revised occlusion interval reference) and image stabilization amplitude are directly retrieved as initialization, i.e., the starting node attributes and motion field subtraction vector of the preset temporal correction map. Frame order correction (a local version of restricted path search) and motion estimation (an initialization version of hierarchical block matching) are performed preferentially in the similar map path (the sub-path of the effective edge set) and spatial band (the geometric region of the frame set), limiting the search depth to, for example, 50% of the original algorithm to reduce the cost. If it is not a hit, it returns to the standard process.

[0089] After scene similarity retrieval and initialization are completed, the processing of new video clips can be started quickly with the help of the mapping index. In continuous detection of road monitoring, the convergence time of temporal correction and motion estimation is significantly shortened, and an efficient reuse path is provided for similar events such as repeated traffic congestion. At the same time, by prioritizing similar paths and spatial bands, the relevance of the initialization values ​​is ensured, the computational overhead is reduced and the continuity and real-time performance of the judgment are improved, and the resource utilization of edge devices in long time series is optimized.

[0090] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0091] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0092] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0093] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0094] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for rapid detection of road video events based on edge computing, characterized in that, Comprising steps of: S1: reading a road video clip, constructing a timing correction graph according to inter-frame content similarity and boundary continuity, generating a corrected sequence by rearranging the arrival order according to the correction graph, and outputting a frame consistency index; S2: estimating a stable motion field and a continuity score on the corrected sequence, writing the image stabilization amplitude back to the timing correction graph to correct the time anchor points of the occlusion interval, and providing a time reference for subsequent evidence generation; Step S2 performs hierarchical block matching on the corrected sequence to estimate a global micro-vibration trajectory in the long static area, subtracts the trajectory from the displacement field to obtain a stable motion field, compensates for the inter-row time difference caused by line-by-line exposure by fitting a linear drift of the background displacement along the row direction, suppresses the rolling shutter deformation, generates a continuity score based on two consistency measures, the direction consistency rate of the same point in a short window, and the sparsity of curvature change along the trajectory, and fuses the two to form a single score to quantify the timing continuity. The image stabilization amplitude calculated from the background block is written back to the timing correction graph to update the time anchor points of the occlusion interval, so that the graph structure and motion estimation share the same time reference; Perform hierarchical block matching on the corrected sequence, divide each frame into fixed-size image blocks, identify the long static area, calculate the inter-block displacement vector using the Lucas-Kanade optical flow method for blocks in the long static area, obtain a global micro-vibration trajectory, and calculate the full-frame displacement field. Subtract the global micro-vibration trajectory from the initial displacement field block by block to obtain a stable motion field; S3: generating an event evidence sequence based on the stable motion field, merging trajectory breakpoints according to the timing correction graph, and initiating a backtracking revision request for the continuity score and returning the updated candidate event clip; S4: performing consistency verification on the candidate event, if there is a conflict with the frame consistency index, revising the timing correction graph and the image stabilization amplitude across levels and triggering stable motion field re-estimation, and outputting the confirmed event after re-evaluation; S5: publishing the trigger result, and mapping the confirmed event, event evidence sequence, and timing correction graph as initialization priors and cross-level recall basis for new clip, to accelerate the convergence of timing correction and judgment of subsequent clips.

2. The edge computing-based road video event rapid detection method according to claim 1, characterized in that: Step S1 takes the video clip as input, calculates the texture histogram similarity and edge band overlap between adjacent frames to obtain two types of observable consistency signals, constructs a timing correction graph with each frame as a node, and the edge weight is based on the joint ordering of content similarity and boundary continuity. Weak consistency edges are directly removed, a restricted shortest path search is performed on the graph, the path constraints are time stamp adjacency and structural continuity, the most coherent frame sequence is obtained as the corrected sequence, the frame consistency index is generated based on the corrected sequence, the index records the front and rear adjacency, available interval, and conflict marker of the frame, forming a dense retrieval entry for timing verification.

3. The edge computing-based road video event rapid detection method according to claim 2, characterized in that: Step S3 extracts three types of atomic evidence from the stable motion field, which are directed to velocity abrupt change, path deviation and space occupancy change, records the appearance time, duration and local geometric relationship of the three types of evidence, calls the adjacency and time anchor information of the timing correction graph, performs shape and texture matching in a small range before and after the occlusion, performs bounded connection for the trajectory breakpoint caused by occlusion, generates an event evidence sequence, triggers a backtracking revision request for a section with a low continuity score, the request contains frame interval, displacement pattern and conflict type, indicates the re-estimation of motion and time anchor in the original section, summarizes the evidence after connection, outputs the candidate event segment, and attaches the corresponding continuity score and evidence list.

4. The edge computing-based road video event rapid detection method according to claim 3, characterized in that: Step S4 verifies the consistency of the candidate event segment, checks the consistency relationship between the frame sequence and the main path on the graph with the frame consistency index, checks whether the motion direction and the geometric change are in the same direction with the continuity score, detects the conflict, performs local relaxation on the timing correction graph, performs limited rearrangement on the local frame sequence according to the joint ordering of content similarity and boundary continuity, adjusts the pan amplitude according to the displacement residual of the background block, after revision, only re-estimates the stable motion field and re-checks the continuity score in the affected interval, aligns and re-evaluates the event evidence sequence with minimum changes, and forms a confirmed event.

5. The edge computing-based road video event rapid detection method according to claim 4, characterized in that: Step S5 generates a trigger result for the confirmed event and publishes it, establishes a mapping index, the index takes the event identifier as the key and maps to the frame set involved in the judgment, the effective edge set of the timing correction graph, the atomic evidence list of the event evidence sequence and the continuity score interval, when a new segment enters the process, performs scene similarity search according to the mapping index, if the search hits, directly calls the corresponding time anchor and pan amplitude as the initialization value, and preferentially performs frame sequence correction and motion estimation in the similar graph path and space band, reducing the reconstruction cost and shortening the convergence time.

6. The edge computing-based road video event rapid detection method according to claim 2, characterized in that: Step S1 extracts the texture distribution vector of each frame between adjacent frames by the local binary pattern histogram method, calculates the texture histogram similarity using the cosine similarity technique, extracts the edge pixel set of each frame using the Canny edge detection, and calculates the edge overlap degree using the Hausdorff distance technique, obtaining two types of observable consistency signals.

7. The edge computing-based road video event rapid detection method according to claim 6, characterized in that: Step S1 constructs a timing correction graph with each frame as a node, for any node pair, the weight of the edge is determined by the joint ordering of content similarity and boundary continuity, the similarity and the overlap degree are normalized to obtain rank values, the joint ordering value is obtained by adding the two rank values, and the reciprocal of the joint ordering value is taken as the edge weight, and the joint ordering value exceeding the preset threshold is determined as a weak consistency edge and is removed. 8.The edge computing based road video event rapid detection method according to claim 1, characterized in that: In step S2, the background region is identified from the stable motion field with the displacement amplitude of the pixel row being lower than the average displacement of the sequence, the background displacement is extracted along the row direction of each frame, the least square method is used to fit the linear drift model for the row-level displacement sequence, the linear drift fitting is applied to each row of pixels for compensation, the stable motion field is updated so that the displacement of all rows is consistent with the compensated value, and the rolling shutter deformation is inhibited.

Images