A cross-frame jump trajectory recovery method, device, equipment and storage medium

Abstract

The application provides a cross-frame jump trajectory recovery method, device and equipment and storage medium, wherein the method first acquires target detection results of a current video frame and historical trajectory information of targets in the video. The historical trajectory information is used to analyze the current frame motion state of the target and determine whether the target has cross-frame jump motion. If it is determined that there is cross-frame jump motion, a two-stage progressive, back-up replacing front-stage failure cascaded trajectory recovery mechanism is started to generate a predicted trajectory. If the target motion is normal, a conventional motion prediction model is used to generate a predicted trajectory. The target detection results and the predicted trajectory are matched to determine the target identity. Finally, the historical trajectory of the corresponding target is updated using real-time detection results according to the target identity obtained by matching, so as to realize continuous tracking of the video target. By using the above method, the continuity and stability of multi-target video tracking under complex working conditions can be significantly improved while ensuring the tracking efficiency in a conventional scene.

Description

Technical Field

[0001] This application relates to the field of computer vision technology, and more specifically, to a method, apparatus, device, and storage medium for recovering cross-frame jump trajectories. Background Technology

[0002] Multi-target video tracking technology is widely used in various scenarios such as infrared small target detection, maritime ship monitoring, low-altitude UAV control, and long-range field security. Due to hardware limitations such as bandwidth, storage, and power consumption of the front-end acquisition equipment, most field and airborne monitoring equipment often use low frame rate mode to acquire video data. However, small targets in the scene generally have high-speed maneuverability, and continuous tracking of small targets under low frame rate conditions is a common technical challenge in the current field.

[0003] The existing SORT tracking algorithm employs a tracking architecture that combines motion prediction with frame matching. It predicts the target's current position based on historical trajectory coordinates using Kalman filtering and relies solely on calculating the intersection-over-union (IoU) ratio between the predicted and real-time detected frames to complete target association and pairing between video frames. This algorithm tracks solely based on the overlap between the target position and the frame, lacking capabilities for multi-dimensional feature matching, motion anomaly detection, and trajectory repair and compensation.

[0004] The algorithm's correlation matching relies entirely on the spatial overlap of detection boxes in adjacent frames. Low frame rate acquisition modes increase the inter-frame time interval, and high-speed small targets are prone to situations such as a sharp increase in inter-frame displacement, no overlap between detection boxes in consecutive frames, and a cross-union ratio (CUP) of zero. The algorithm lacks corresponding anomaly identification and failure compensation mechanisms. When the CUP matching fails, it cannot complete inter-frame target correlation, directly leading to broken target tracking trajectories and target tracking loss. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, apparatus, device and storage medium for cross-frame jump trajectory recovery, which can significantly improve the continuity and stability of multi-target video tracking under complex conditions while ensuring tracking efficiency in conventional scenarios.

[0006] In a first aspect, embodiments of this application provide a method for recovering cross-frame jump trajectories, the method comprising: Obtain the target detection results in the current video frame; Obtain historical trajectory information of each target that has appeared in the video; Based on the historical trajectory information, anomaly analysis is performed on the motion state of each target in the current frame to determine whether cross-frame jump motion has occurred. When a target is determined to have a cross-frame jump motion, a cascaded trajectory recovery mechanism is triggered to generate the predicted trajectory of each target. The cascaded trajectory recovery mechanism includes at least two recovery stages executed in a preset order, wherein the later stage is triggered when the recovery of the previous stage fails. If no cross-frame jump motion is determined, the predicted trajectories of the targets are generated based on the conventional motion prediction model; The target detection results are matched with the predicted trajectories of each target to determine the target identity corresponding to each target detection result; Based on the determined target identity, the historical trajectory information of the corresponding target is updated using the target detection results to maintain continuous tracking of the target.

[0007] Optionally, determining whether a cross-frame jump motion has occurred includes determining that a cross-frame jump motion has occurred in response to one or more of the following conditions being met: The target's displacement suddenly increases beyond a preset displacement threshold in N consecutive frames; The Intersection over Union (IoU) between the target and the detection results remains zero in the current frame and in consecutive preceding frames; The target maintains a consistent direction of motion across multiple consecutive frames and its rate of velocity change is below a preset rate of change threshold. The change in the target's detection bounding box size is less than a preset size threshold.

[0008] Optionally, the cascaded trajectory recovery mechanism includes: The first stage involves predicting and recovering the trajectory based on motion information from historical trajectories, generating a preliminary predicted trajectory, and determining the initial position of the target in the current frame based on the preliminary predicted trajectory. If the preliminary predicted trajectory cannot establish a valid match with any target detection result, the first stage recovery is deemed to have failed. Second stage: When the recovery of the first stage fails, a local image patch multi-scale search is performed in the vicinity of the initial position, and the initial position is corrected according to the search matching results to obtain the final predicted trajectory.

[0009] Optionally, the trajectory prediction and recovery based on historical trajectory motion information includes: Extract the target's location sequence from the historical trajectory information; A motion estimation model is constructed based on the position sequence to calculate the target's motion speed and motion trend; Based on the motion speed and motion trend, the initial position of the target in the current frame is predicted using interpolation or filtering methods.

[0010] Optionally, the local image patch multi-scale search includes: Extract the target's appearance features from historical trajectory information as a matching template; Construct a multi-scale image pyramid for the current frame; On each scale layer of the image pyramid, candidate search regions are delineated based on the initial position; Within the candidate search area, the similarity between each candidate window and the matching template is calculated using a sliding window method; The location of the candidate window with the highest similarity is selected as the corrected target location.

[0011] Optionally, it also includes: The final predicted trajectory is verified for consistency of motion direction and trajectory curvature constraint. If the verification passes, it is output as the predicted trajectory of each target. If the verification fails, the predicted trajectory is discarded and marked as recovery failure. It also includes trajectory lifecycle management: For target detection results that do not match any existing trajectory, initialize new historical trajectory information for them; for existing trajectories that do not match any detection results for multiple consecutive frames, remove them from the tracking queue.

[0012] Optionally, matching the target detection results with the predicted trajectories of each target includes: Calculate the intersection-union ratio, center point distance, and appearance similarity between each detection box and each predicted trajectory; The intersection-union ratio, the center point distance, and the appearance similarity are fused to obtain a comprehensive matching score; Based on the comprehensive matching score, a one-to-one correspondence between the detection results and the predicted trajectory is determined by a bipartite graph matching algorithm, and an identity is assigned to the target.

[0013] Secondly, embodiments of this application provide a cross-frame jump trajectory recovery device, the device comprising: The detection and acquisition module is used to acquire the target detection results in the current video frame and the historical trajectory information of each target that has appeared in the video. The abnormal motion determination module is used to perform anomaly analysis on the motion state of each target in the current frame based on the historical trajectory information, and to determine whether cross-frame jump motion has occurred. The cascaded trajectory recovery module is used to trigger a cascaded trajectory recovery mechanism to generate predicted trajectories for each target when it is determined that the target has undergone cross-frame jump motion. The cascaded trajectory recovery mechanism includes at least two recovery stages executed in a preset order, wherein the latter stage is triggered when the former stage fails to recover. It is also used to generate predicted trajectories for each target based on a conventional motion prediction model if it is not determined that cross-frame jump motion has occurred. The matching and identity determination module is used to match the target detection results with the predicted trajectories of each target to determine the target identity corresponding to each target detection result; The trajectory update module is used to update the historical trajectory information of the corresponding target based on the determined target identity using the target detection results, so as to maintain continuous tracking of the target.

[0014] Optionally, the abnormal motion determination module is specifically used to determine that the target has undergone cross-frame jump motion in response to one or more of the following conditions: the displacement of the target suddenly increases by more than a preset displacement threshold in N consecutive frames; the intersection-union ratio (IoU) of the target with the detection result is continuously zero in the current frame and previous consecutive frames; the target maintains the same motion direction in multiple consecutive frames and the rate of change of velocity is lower than a preset rate of change threshold; the change in the size of the target's detection box is less than a preset size threshold.

[0015] Optionally, the cascaded trajectory recovery mechanism includes a first recovery stage and a second recovery stage; The first recovery stage is used to perform trajectory prediction and recovery based on the motion information of historical trajectories, generate a preliminary predicted trajectory, and determine the preliminary position of the target in the current frame based on the preliminary predicted trajectory; if the preliminary predicted trajectory cannot establish a valid match with any target detection result, the first stage recovery is deemed to have failed. The second recovery stage is used to perform a local image patch multi-scale search in the vicinity of the initial position when the first stage recovery fails, and to correct the initial position according to the search matching results to obtain the final predicted trajectory.

[0016] Optionally, the first recovery stage is specifically used for: extracting the target's position sequence from the historical trajectory information; constructing a motion estimation model based on the position sequence to calculate the target's motion speed and motion trend; and predicting the target's initial position in the current frame using interpolation or filtering methods based on the motion speed and motion trend.

[0017] Optionally, the second recovery stage is specifically used for: extracting the appearance features of the target from historical trajectory information as a matching template; constructing a multi-scale image pyramid for the current frame; delineating candidate search regions on each scale layer of the image pyramid based on the initial position; calculating the similarity between each candidate window and the matching template within the candidate search region using a sliding window method; and selecting the position of the candidate window with the highest similarity as the corrected target position.

[0018] Optionally, the cascaded trajectory recovery module is further configured to: perform motion direction consistency verification and trajectory curvature constraint verification on the final predicted trajectory; if the verification passes, the predicted trajectory is output as the predicted trajectory of each target; if the verification fails, the predicted trajectory is discarded and marked as recovery failure. The device also includes a trajectory lifecycle management module; The trajectory lifecycle management module is used to initialize new historical trajectory information for target detection results that do not match any existing trajectory; and to remove existing trajectories that do not match any detection results for multiple consecutive frames from the tracking queue.

[0019] Optionally, the matching and identity determination module is specifically used to: calculate the intersection-union ratio, center point distance, and appearance similarity between each detection box and each predicted trajectory; fuse the intersection-union ratio, the center point distance, and the appearance similarity to obtain a comprehensive matching score; and based on the comprehensive matching score, determine the one-to-one correspondence between the detection results and the predicted trajectory using a bipartite graph matching algorithm to assign an identity to the target.

[0020] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the cross-frame jump trajectory recovery method described in any of the optional embodiments of the first aspect are performed.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the cross-frame jump trajectory recovery method described in any of the optional embodiments of the first aspect.

[0022] The technical solution provided in this application includes, but is not limited to, the following beneficial effects: By acquiring the target detection results in the current video frame, it is possible to collect effective detection data of all targets in the video frame in real time and completely, providing reliable and comprehensive raw data support for the entire tracking process, including target motion state analysis, trajectory prediction, target matching, and trajectory update, and ensuring the integrity of the data foundation of the overall tracking process.

[0023] By acquiring the historical trajectory information of each target that has appeared in the video, the past motion correlation data of the target can be completely preserved. This provides sufficient historical reference for subsequent target motion anomaly identification and trajectory prediction, avoiding the problem of insufficient accuracy caused by relying solely on single-frame data for calculation and judgment, and improving the credibility of target motion state analysis.

[0024] By analyzing the target's motion state based on historical trajectory information and determining whether there is cross-frame jump motion, the normal motion state of the target and the abnormal cross-frame jump motion state can be accurately distinguished, thus achieving effective identification of the target's special motion conditions and providing accurate pre-judgment basis for the subsequent switching and execution of differentiated trajectory generation strategies.

[0025] By triggering a multi-level progressive cascaded trajectory recovery mechanism to generate a predicted trajectory when the target undergoes cross-frame jump motion, targeted compensation processing can be carried out for conventional tracking failure scenarios caused by cross-frame jumps. This effectively solves the problem of trajectory not being able to be predicted normally under large displacement cross-frame motion of the target and avoids tracking trajectory interruption under abnormal working conditions.

[0026] By generating predicted trajectories using a conventional motion prediction model when no cross-frame jump motion is detected, the computational logic can be simplified in conventional scenarios where the target moves smoothly. There is no need to start a complex trajectory recovery process. While ensuring conventional tracking accuracy, the overall computational overhead of the algorithm is effectively reduced, and the tracking efficiency in conventional scenarios is improved.

[0027] By matching the target detection results with the predicted trajectory and determining the target identity, a precise one-to-one correspondence between the predicted trajectory and the real-time detected target can be achieved, effectively distinguishing multiple different targets in the image, avoiding target confusion and identity misidentification during multi-target tracking, and ensuring the uniqueness and orderliness of multi-target tracking.

[0028] By updating the historical trajectory information of the corresponding target in real time according to the determined target identity, dynamic iteration and continuous optimization of the target trajectory data can be achieved, so that the trajectory data can be adapted to the latest motion state of the target in real time, ensuring the continuity of the long-term tracking process of the target and maintaining a stable and continuous tracking effect.

[0029] In summary, this application, through a complete and interconnected step-by-step tracking logic, can adaptively adapt to both normal and stable target motion and abnormal cross-frame jumping motion, achieving accurate prediction and continuous updating of target trajectory under different motion states. It effectively improves the problems of target trajectory breakage and tracking loss in cross-frame large displacement scenarios, and significantly enhances the continuity and stability of multi-target video tracking in complex scenarios while ensuring tracking efficiency in normal scenarios.

[0030] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of a cross-frame jump trajectory recovery method provided in Embodiment 1 of this application is shown; Figure 2 A flowchart of a trajectory prediction and recovery method provided in Embodiment 1 of this application is shown; Figure 3 A flowchart of a local image patch multi-scale search method provided in Embodiment 1 of this application is shown; Figure 4 A flowchart of a trajectory prediction matching method provided in Embodiment 1 of this application is shown; Figure 5 This paper presents a schematic diagram of the overall process of a cross-frame jump trajectory recovery method provided in Embodiment 1 of this application; Figure 6 This paper shows a schematic diagram of the structure of a cross-frame jump trajectory recovery device provided in Embodiment 2 of this application; Figure 7 A schematic diagram of the structure of a computer device provided in Embodiment 3 of this application is shown. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] Example 1 To facilitate understanding of this application, the following is combined with... Figure 1 The flowchart of the cross-frame jump trajectory recovery method provided in Embodiment 1 of this application illustrates Embodiment 1 of this application in detail.

[0035] See Figure 1As shown, Figure 1 A flowchart of a cross-frame jump trajectory recovery method provided in Embodiment 1 of this application is shown, wherein the method includes steps S101 to S107: S101: Obtain the target detection results in the current video frame.

[0036] Specifically, after reading low frame rate video footage, the data is sent to the target detection module. The module outputs the detection bounding box data for all targets within the frame. One path of this detection data is used for regular IoU (Intersection over Union) correlation calculations, while the other path is reserved for subsequent trajectory recovery. This entire method can be applied to infrared small target tracking, UAV target tracking, sea surface target tracking, long-range surveillance, low frame rate video analysis, and weak texture small target tracking scenarios.

[0037] Currently, mainstream tracking algorithms such as SORT (Simple Online and Realtime), DeepSORT (Deep Simple Online and Realtime Tracking, an enhanced online real-time tracking algorithm that integrates deep appearance features), ByteTrack (Byte based Tracking, a tracking algorithm based on hierarchical association of high and low confidence detection boxes), and OC-SORT (Observation-Centric SORT, an improved SORT tracking algorithm centered on observation) all use detection results as algorithm input, but they only rely on detection boxes to complete IoU matching or feature matching, and do not have a corresponding cross-frame broken trajectory repair process.

[0038] Furthermore, the program continuously reads the bitstream data through the system's built-in video decoding interface, splits the image into single frames at a fixed frame rate, and sends the image pixel matrix into the detection and inference module. After the detection module finishes its operation, it encapsulates structured data such as target coordinates, bounding box dimensions, and confidence scores, and divides the memory into two independent cache blocks to store regular matching data and backup recovery data, respectively.

[0039] S102: Obtain the historical trajectory information of each target that has appeared in the video.

[0040] Specifically, the system stores trajectory data separately for each tracked target. The stored content includes the target's historical frame position coordinates, frame-by-frame motion direction, instantaneous velocity, and trajectory curvature. At the same time, it archives the target's local image patches cropped from the historical frames. All of the above stored content is essential data for subsequent abnormal motion judgment, hierarchical trajectory recovery, and image matching template creation.

[0041] Traditional algorithms such as SORT series and ByteTrack only cache a few frame coordinates for a short time, and only satisfy the basic operations of Kalman filtering. They do not save trajectory curvature and target image data, and cannot realize multi-level trajectory repair logic.

[0042] Furthermore, the system predefines the trajectory structure data format, and each trajectory is automatically assigned a unique ID (Identity, unique identifier of the target). The program uses the ID as an index key to address the corresponding memory address and writes various trajectory parameters and cropped images sequentially according to the frame time sequence.

[0043] S103: Based on the historical trajectory information, perform anomaly analysis on the motion state of each target in the current frame to determine whether cross-frame jump motion has occurred.

[0044] Specifically, the algorithm is executed in a three-stage logical sequence. It prioritizes the regular IoU association matching operation. Only when the target fails to match IoU for multiple consecutive frames will it retrieve the target's full historical trajectory data and comprehensively determine whether the target has jumped across frames based on four judgment conditions.

[0045] This anomaly detection logic differs from the traditional algorithm's approach of determining target loss based solely on missed detections. Traditional algorithms assume that the target displacement in adjacent frames is small and that the detection boxes overlap. When the frame rate is low and the high-speed movement of small targets causes large displacements between frames, they cannot identify cross-frame jumps, directly resulting in data association failure and broken tracking trajectories.

[0046] Furthermore, the program iterates through all targets, sets a counter to count the number of consecutive IoU failure frames for a single target, and after the counter reaches the target, it retrieves the displacement, direction, and size parameters in sequence and compares them with the configured thresholds one by one.

[0047] S104: When it is determined that the target has a cross-frame jump motion, a cascaded trajectory recovery mechanism is triggered to generate the predicted trajectory of each target; the cascaded trajectory recovery mechanism includes at least two recovery stages executed in a preset order, wherein the latter stage is triggered when the former stage fails to recover.

[0048] Specifically, trajectory recovery is divided into two-stage processing steps. The first stage relies on historical motion parameters to predict and recover the trajectory. If the predicted position output in this stage cannot form an effective match with any detection box, the recovery is deemed to have failed, and the program automatically switches to the second stage, which uses local image block pyramid search to complete the position correction. The two-stage recovery plus the pre-processed conventional IoU association together form a complete three-layer processing architecture.

[0049] Existing tracking algorithms lack a hierarchical trajectory recovery design. When IoU matching fails, the only options are to extend the trajectory buffer time, perform simple Kalman prediction, or directly delete the entire failed trajectory. There is a lack of dedicated recovery methods for cross-frame jump scenarios.

[0050] Furthermore, a Boolean flag is set in the program to control branch jumps. After the first-level recovery operation is completed, all detection boxes are traversed in a loop to perform position matching. If no matching result is found, the flag is changed and the program jumps to start the second-level image search process.

[0051] S105: If no cross-frame jump motion is determined, the predicted trajectory of each target is generated based on the conventional motion prediction model.

[0052] Specifically, the available motion prediction models include five categories: Kalman filtering, uniform motion models, multinomial trajectory fitting models, curve trajectory prediction models, and deep learning-based trajectory prediction models. Under normal conditions where the target motion is stable and there are no frame jumps, any one of these models can be directly selected to calculate the predicted position, without initiating image pyramid construction and image matching related operations, and using mature traditional prediction logic.

[0053] Furthermore, model selection parameters are reserved in the configuration file, and the program reads the parameters to load the corresponding model. Under normal operating conditions, the function call instructions for pyramid initialization and sliding matching are blocked.

[0054] S106: Match the target detection results with the predicted trajectories of each target to determine the target identity corresponding to each target detection result.

[0055] Specifically, traditional matching schemes are divided into two types: single IoU overlap matching and single appearance ReID (Re-Identification) feature matching. Infrared targets and small targets at long distances have few pixels and lack texture, and environmental noise can easily interfere with the stability of appearance features, easily leading to false associations and frequent switching of target identities. This scheme integrates multi-dimensional parameters to carry out comprehensive matching, which can effectively reduce the probability of small target missed tracking, trajectory drift, and false matching.

[0056] Furthermore, a two-layer for loop is used to implement full combination traversal. The outer loop traverses all trajectories and detection boxes, and sequentially completes the numerical calculation of the three types of indicators. The calculation results are stored in a temporary array to await weighted fusion.

[0057] S107: Based on the determined target identity, update the historical trajectory information of the corresponding target using the target detection results to maintain continuous tracking of the target.

[0058] Specifically, after identity matching is completed, the target coordinates, detection box size, and captured local image of the target in the current frame are added to the historical trajectory storage space of the corresponding target one by one. The continuously updated trajectory data will continue to participate in the motion anomaly judgment and trajectory prediction calculation in the next frame, and the long-term uninterrupted tracking of small targets can be achieved by relying on the continuously accumulated historical data.

[0059] Furthermore, based on the target ID, the trajectory storage address is located, and various data are updated by appending to it. The write operation is performed immediately after the single-frame matching is completed.

[0060] In an optional implementation, determining whether a cross-frame jump motion has occurred includes determining that a cross-frame jump motion has occurred in response to one or more of the following conditions being met: The target's displacement suddenly increases beyond a preset displacement threshold in N consecutive frames.

[0061] Specifically, the displacement threshold and the number of consecutive frames N can be flexibly adjusted to suit different application scenarios such as sea surface monitoring, drone inspection, and infrared security. The program calculates the displacement distance of the target's center point frame by frame, and accumulates the total displacement of N consecutive frames. If the value exceeds the set threshold, it is determined that the target has generated cross-frame jump motion, thereby capturing the motion characteristics of the target that change drastically in low frame rate environments.

[0062] Furthermore, N and the critical displacement values ​​are stored as configuration parameters. The program uses the Euclidean distance formula to calculate the inter-frame center point spacing and uses an accumulation command to calculate the displacement of N consecutive frames.

[0063] The Intersection over Union (IoU) between the target and the detection results remains zero in the current frame and in the preceding consecutive frames.

[0064] Specifically, the IoU values ​​between the target prediction box and all detection boxes in the frame are calculated one by one. When there is no overlap and the IoU is always zero for multiple consecutive frames, it means that the traditional IoU association has completely failed. This condition serves as the key criterion for starting cascaded trajectory recovery.

[0065] Furthermore, the overlap ratio is calculated in batches using the standard IoU calculation formula, and a frame counter is set to record the number of consecutive zero IoUs. Once the count value reaches the target, the recovery process is triggered.

[0066] The target maintains a consistent direction of motion across multiple consecutive frames, and its rate of change of velocity is below a preset rate of change threshold.

[0067] Specifically, the target's direction of travel and speed are extracted from historical trajectory data. When the amplitude of directional fluctuations and speed changes are less than a set threshold for multiple consecutive frames, the object's motion continuity is combined with the law of motion to determine a cross-frame jump, thus avoiding erroneous triggering caused by random noise in the image.

[0068] Furthermore, the motion azimuth angle and instantaneous velocity are calculated from the coordinate difference and compared with preset thresholds. If both conditions are met simultaneously, the judgment rule is met.

[0069] The change in the target's detection bounding box size is less than a preset size threshold.

[0070] Specifically, by comparing the current detection box size with the average size of historical multi-frame boxes, when the size change is less than a set threshold, the system is combined with the other three conditions for joint verification to distinguish between two scenarios: the original target moving across frames and a completely new target entering the frame, thereby reducing errors in anomaly detection.

[0071] Furthermore, the average value of historical multi-frame width and height data is calculated, and the change ratio of the current frame relative to the average value is calculated. The ratio data is compared with the threshold and used as an auxiliary judgment parameter.

[0072] In an optional implementation, the cascaded trajectory recovery mechanism includes: The first stage involves predicting and recovering the trajectory based on motion information from historical trajectories, generating a preliminary predicted trajectory, and determining the initial position of the target in the current frame based on the preliminary predicted trajectory. If the preliminary predicted trajectory cannot establish a valid match with any target detection result, the first stage recovery is deemed to have failed.

[0073] Specifically, the system extracts information on the direction of motion, speed, trajectory curvature, and trajectory change trends from historical trajectories, and uses various selectable motion prediction models to calculate the initial coordinates of the target in the current frame. If no matching target box is found after traversing all detection boxes, the marking stage recovery fails, and the system automatically jumps to the next stage of the image search process.

[0074] Furthermore, four motion parameters are read in batches from the trajectory storage structure and imported into the selected prediction model as input parameters. After calculating and outputting the coordinates, the entire detection box is traversed for spatial comparison. If no matching item is found, the failure flag is set.

[0075] Second stage: When the recovery of the first stage fails, a local image patch multi-scale search is performed in the vicinity of the initial position, and the initial position is corrected according to the search matching results to obtain the final predicted trajectory.

[0076] Specifically, the local candidate retrieval region is delineated centered on the preliminary coordinates output in the first stage. Matching retrieval is carried out based on the multi-scale image pyramid structure and historically retained target image templates. The predicted coordinates are corrected using the best matching results, thereby making up for the matching shortcomings caused by the lack of texture and the failure of appearance features of small targets.

[0077] Furthermore, the configuration pixel augmentation parameters are read, and a rectangular search range is generated by extending outwards from the reference coordinates. Template retrieval is then performed within the same rectangular area of ​​each pyramid layer image.

[0078] In an optional implementation, see Figure 2 As shown, Figure 2The flowchart of a trajectory prediction and recovery method provided in Embodiment 1 of this application is shown, wherein the trajectory prediction and recovery based on motion information of historical trajectories includes steps S201 to S203: S201: Extract the target's location sequence from the historical trajectory information.

[0079] Specifically, according to the chronological order of the video, the historical coordinate data of the target throughout the entire time period are retrieved, arranged and spliced ​​in an orderly manner to form a complete position sequence, which serves as the raw input data for the motion estimation model.

[0080] Furthermore, all historical coordinates are sorted from earliest to latest based on the frame timestamp, outlier points are removed, and the data is formatted into a one-dimensional array.

[0081] S202: Construct a motion estimation model based on the position sequence to calculate the target's motion speed and motion trend.

[0082] Specifically, any preset motion model is selected, and calculations are performed based on an ordered sequence of positions to obtain key motion parameters such as the instantaneous velocity, direction of travel, and trajectory curvature of the target.

[0083] Furthermore, the instantaneous velocity is obtained through adjacent coordinate difference calculation, the trajectory curvature is solved by multi-point fitting, all calculation parameters are temporarily cached, and directly supplied to the position prediction stage.

[0084] S203: Based on the motion speed and motion trend, predict the initial position of the target in the current frame using interpolation or filtering methods.

[0085] Specifically, Kalman filtering algorithms are selected, and multiple constraints such as velocity, motion trend, and trajectory curvature are combined to complete the position deduction, suppress the drift of the predicted point, and output the reference coordinates that can be used for subsequent image search.

[0086] Furthermore, velocity, direction, and curvature are used as constraints in the filtering iteration formula to output the coordinates of a single pixel.

[0087] In an optional implementation, see Figure 3 As shown, Figure 3 The flowchart of a local image patch multi-scale search method provided in Embodiment 1 of this application is shown, wherein the local image patch multi-scale search includes steps S301 to S305: S301: Extract the appearance features of the target from historical trajectory information as a matching template.

[0088] Specifically, target image blocks with intact and unobstructed images from historical frames are selected to create matching templates, avoiding the problem of poor image quality and insufficient effective texture for small targets in the current frame, which prevents the generation of reliable templates.

[0089] Furthermore, the system iterates through historical archived images, selects the samples with the best clarity, and uniformly scales them to a fixed size before storing them in template memory.

[0090] S302: Construct the multi-scale image pyramid for the current frame.

[0091] Specifically, the current whole frame image is sampled at multiple scaling ratios to generate multi-layer image pyramids with different resolutions, adapting to the retrieval needs of infrared and long-distance small targets with different imaging sizes.

[0092] Furthermore, using the original image as the top layer, multiple layers of images are generated by downsampling layer by layer according to a preset scaling factor. The number of pyramid layers and the scaling factor can be configured and adjusted.

[0093] S303: On each scale layer of the image pyramid, candidate search regions are delineated based on the initial position.

[0094] Specifically, a fixed range is extended outward from the initial predicted coordinates to define a local rectangular search area. This search boundary is used in each layer of the pyramid image, and matching calculations are performed only within the defined area, saving the unnecessary computation caused by traversing the entire image.

[0095] Furthermore, the horizontal and vertical expansion pixel parameters are read and configured, and the same search box coordinates are reused in each pyramid layer.

[0096] S304: Within the candidate search area, calculate the similarity between each candidate window and the matching template using a sliding window method.

[0097] Specifically, the sliding window traverses the candidate regions with a fixed step size. The similarity calculation can switch between five algorithms as needed: NCC (Normalized Cross-Correlation) template matching, optical flow matching, feature point matching, Transformer (Transformer feature extraction model) similarity search, and Siamese (Siamese Network) network matching. The matching scores of each window and template are calculated one by one.

[0098] Furthermore, the step size parameter is configurable, and the five matching algorithms are modular and pluggable, allowing for switching of the corresponding similarity calculation scheme according to different imaging scenarios such as visible light and infrared.

[0099] S305: Select the location of the candidate window with the highest similarity as the corrected target location.

[0100] Specifically, the similarity scores of all windows within the selected area are compared, and the coordinates of the window with the highest score are used to replace the initial predicted point to complete the target position correction.

[0101] Furthermore, by traversing all the stored similarity data to complete the extreme value filtering, the optimal coordinates directly overwrite the original preliminary coordinates.

[0102] In an optional implementation, it also includes: The final predicted trajectory is verified for consistency of motion direction and trajectory curvature constraint. If the verification passes, it is output as the predicted trajectory of each target. If the verification fails, the predicted trajectory is discarded and marked as recovery failure.

[0103] Specifically, by referring to the inherent directional change range and curvature fluctuation range of the target's historical trajectory, and comparing it with the corrected trajectory parameters, if the parameters exceed the reasonable range, the predicted trajectory is directly discarded and marked as recovery failure, thereby reducing false trajectories generated by incorrect matching.

[0104] Furthermore, retrieve the upper and lower limits of historical parameters and compare the direction and curvature respectively. If any one of them exceeds the limit, the trajectory is invalidated.

[0105] For target detection results that do not match any existing trajectories, initialize new historical trajectory information for them.

[0106] Specifically, a brand new trajectory file is created for each newly appearing target in the image, and information such as the detection coordinates of the current frame and the appearance image of the target is recorded. The new trajectory is then incorporated into the global tracking queue to participate in the calculation of subsequent frames.

[0107] Furthermore, the system automatically generates a brand new ID and allocates an independent storage space to store the new trajectory data. The new trajectory will participate in anomaly detection and trajectory prediction normally starting from the next frame.

[0108] For existing trajectories that have not been matched with any detection results for multiple consecutive frames, remove them from the tracking queue.

[0109] Specifically, the matching status of each trajectory is statistically analyzed in real time. When a trajectory cannot find a corresponding detection box for multiple consecutive frames, the trajectory is deleted from the tracking queue and invalid stored data is cleaned up.

[0110] Furthermore, each trajectory is equipped with an independent mismatch counter. Once the number of consecutive mismatch frames reaches a preset threshold, a deletion operation is performed, and memory is released simultaneously.

[0111] In an optional implementation, see Figure 4 As shown, Figure 4 The flowchart of a predicted trajectory matching method provided in Embodiment 1 of this application is shown, wherein matching the target detection result with the predicted trajectory of each target includes steps S401 to S403: S401: Calculate the intersection-union ratio, center point distance, and appearance similarity between each detection box and each predicted trajectory.

[0112] Specifically, it iterates through all combinations of predicted trajectories and detection boxes, and independently calculates three types of indicators: IoU value, Euclidean distance of center point, and image appearance similarity, to make up for the defect that single IoU directly fails in cross-frame scenarios without overlapping boxes.

[0113] Furthermore, a double loop is used to complete the full combination calculation, and the three results are stored separately in a two-dimensional array.

[0114] S402: The intersection-union ratio, the center point distance, and the appearance similarity are fused to obtain a comprehensive matching score.

[0115] Specifically, the three indicators are weighted and fused to generate a unified matching score, which reduces the matching error caused by image noise distorting the appearance features of small targets and the inaccuracy of single-point indicators.

[0116] Furthermore, the weighting coefficients are preset in advance, and after weighting, they are uniformly normalized to form standardized scores.

[0117] S403: Based on the comprehensive matching score, the one-to-one correspondence between the detection results and the predicted trajectory is determined by the bipartite graph matching algorithm, and an identity is assigned to the target.

[0118] Specifically, a bipartite graph is constructed using the comprehensive matching score as the weight. The optimal matching of the bipartite graph is used to complete the pairing of detection boxes and trajectories, fix the target ID, and effectively suppress the problem of frequent ID switching during the tracking process.

[0119] Furthermore, one-to-one binding is achieved based on the optimal allocation rule, and the original target identifier is used after successful pairing.

[0120] See Figure 5 As shown, Figure 5 This diagram illustrates the overall flow of a cross-frame jump trajectory recovery method provided in Embodiment 1 of this application. The flowchart fully embodies the three-level segmented processing logic of cross-frame jump trajectory recovery in this application, and corresponds one-to-one with the technical steps described in S101-S107 of the specification. The step-by-step flow is explained below: 1. The process begins with video input, corresponding to the video source reading step in S101. The video data is sent to the target detection stage to obtain the target detection result of the current frame and output the target detection box data as the original data source for the entire process. After the detection is completed, the first stage of annotation, the conventional IoU association matching, is entered. This is the basic matching logic that is executed first by default in this solution. The initial IoU association pairing is completed based on the predicted boxes of historical trajectories and the real-time detection boxes.

[0121] 2. After the first stage of matching is completed, proceed to the branch to determine whether the node IOU has failed consecutively: (1) If the judgment result is negative, it means that the target has no cross-frame jump motion. There is no need to start the trajectory recovery process. The process will jump directly to the end result. This scenario corresponds to S105. The predicted trajectory is generated based on the conventional motion prediction model, and the tracking result is output after multi-dimensional matching is completed. (2) If the judgment result is yes, that is, the IoU continuous matching fails, the program enters the second stage of the annotation processing link and executes abnormal motion analysis, triggering trajectory recovery mode, and trajectory prediction recovery in sequence; among them, abnormal motion analysis corresponds to the cross-frame jump judgment logic of S103. Based on the historical trajectory, it comprehensively judges whether the target has jumped across frames from four dimensions: displacement, IoU, motion direction, and detection box size. After determining the cross-frame jump, the cascaded trajectory recovery mechanism is triggered. Trajectory prediction recovery is the first stage processing content of cascaded recovery. The initial position of the target is obtained by calculating based on historical motion parameters.

[0122] 3. After the trajectory prediction and recovery calculation is completed, proceed to the decision node for whether to restore the trajectory: (1) If the determination is yes, that is, the preliminary predicted trajectory can be effectively matched with the existing detection box, the result is directly transferred, and the tracking data is output after identity matching; (2) If the determination is negative, that is, the first stage trajectory prediction recovery fails and cannot match any detection box, the process enters the third stage of the bottom recovery link, and executes the local image pyramid search, direction continuity verification, recovery of target trajectory, and trajectory update in sequence: the local image pyramid search corresponds to the multi-scale image retrieval implementation steps of the second stage of cascade recovery, constructs an image pyramid based on the initial position and slides the template to match and correct the coordinates; the direction continuity verification corresponds to the trajectory validity verification rules of motion direction consistency + trajectory curvature constraint in the specification, and filters out false trajectories generated by incorrect matching; after the verification is qualified, the target trajectory is recovered to obtain the final predicted trajectory; the final trajectory update corresponds to S107, which uses the detection data of the current frame to update the historical trajectory information of the target.

[0123] 4. After all the steps in the third stage are completed, the results are summarized at the bottom of the flowchart, and the final target tracking results are output.

[0124] Overall, the second stage of trajectory prediction and recovery and the third stage of local image pyramid search in the flowchart correspond to the two-level recovery stages executed sequentially by the cascaded trajectory recovery mechanism of this application. The direction continuity verification and trajectory update correspond to the trajectory verification and historical trajectory maintenance content of the scheme, respectively, and fully cover all the technical execution logic of this application.

[0125] Example 2 See Figure 6 As shown, Figure 6 A schematic diagram of a cross-frame jump trajectory recovery device provided in Embodiment 2 of this application is shown, wherein the device includes: The detection and acquisition module 601 is used to acquire the target detection results in the current video frame and the historical trajectory information of each target that has appeared in the video. The abnormal motion determination module 602 is used to perform anomaly analysis on the motion state of each target in the current frame based on the historical trajectory information, and to determine whether cross-frame jump motion has occurred. The cascaded trajectory recovery module 603 is used to trigger a cascaded trajectory recovery mechanism to generate predicted trajectories for each target when it is determined that the target has undergone cross-frame jump motion; the cascaded trajectory recovery mechanism includes at least two recovery stages executed in a preset order, wherein the latter stage is triggered when the former stage fails to recover; and is used to generate predicted trajectories for each target based on a conventional motion prediction model if it is not determined that cross-frame jump motion has occurred. The matching and identity determination module 604 is used to match the target detection results with the predicted trajectories of each target to determine the target identity corresponding to each target detection result; The trajectory update module 605 is used to update the historical trajectory information of the corresponding target based on the determined target identity using the target detection result, so as to maintain continuous tracking of the target.

[0126] In an optional implementation, the abnormal motion determination module is specifically used to determine that the target has undergone cross-frame jump motion in response to one or more of the following conditions: the displacement of the target suddenly increases by more than a preset displacement threshold in N consecutive frames; the intersection-union ratio (IoU) of the target with the detection result is continuously zero in the current frame and previous consecutive frames; the target maintains a consistent motion direction and the rate of change of velocity is lower than a preset rate of change threshold in multiple consecutive frames; and the change in the size of the target's detection box is less than a preset size threshold.

[0127] In one optional implementation, the cascaded trajectory recovery mechanism includes a first recovery phase and a second recovery phase; The first recovery stage is used to perform trajectory prediction and recovery based on the motion information of historical trajectories, generate a preliminary predicted trajectory, and determine the preliminary position of the target in the current frame based on the preliminary predicted trajectory; if the preliminary predicted trajectory cannot establish a valid match with any target detection result, the first stage recovery is deemed to have failed. The second recovery stage is used to perform a local image patch multi-scale search in the vicinity of the initial position when the first stage recovery fails, and to correct the initial position according to the search matching results to obtain the final predicted trajectory.

[0128] In an optional implementation, the first recovery stage is specifically used for: extracting the target's position sequence from the historical trajectory information; constructing a motion estimation model based on the position sequence to calculate the target's motion speed and motion trend; and predicting the target's initial position in the current frame using interpolation or filtering methods based on the motion speed and motion trend.

[0129] In an optional implementation, the second recovery stage is specifically used for: extracting the appearance features of the target from historical trajectory information as a matching template; constructing a multi-scale image pyramid for the current frame; delineating candidate search regions on each scale layer of the image pyramid based on the initial position; calculating the similarity between each candidate window and the matching template within the candidate search region using a sliding window method; and selecting the position of the candidate window with the highest similarity as the corrected target position.

[0130] In an optional implementation, the cascaded trajectory recovery module is further configured to: perform motion direction consistency verification and trajectory curvature constraint verification on the final predicted trajectory; if the verification passes, the predicted trajectory is output as the predicted trajectory of each target; if the verification fails, the predicted trajectory is discarded and marked as recovery failure. The device also includes a trajectory lifecycle management module; The trajectory lifecycle management module is used to initialize new historical trajectory information for target detection results that do not match any existing trajectory; and to remove existing trajectories that do not match any detection results for multiple consecutive frames from the tracking queue.

[0131] In an optional implementation, the matching and identity determination module is specifically used to: calculate the intersection-union ratio, centroid distance, and appearance similarity between each detection box and each predicted trajectory; fuse the intersection-union ratio, centroid distance, and appearance similarity to obtain a comprehensive matching score; and, based on the comprehensive matching score, determine the one-to-one correspondence between the detection results and the predicted trajectories using a bipartite graph matching algorithm to assign an identity to the target.

[0132] Example 3 Based on the same application concept, see [link / reference] Figure 7 As shown, Figure 7 This illustration shows a structural schematic diagram of a computer device provided in Embodiment 3 of this application, wherein, as shown... Figure 7 As shown, the computer device 700 provided in Embodiment 3 of this application includes: The computer device 700 includes a processor 701, a memory 702, and a bus 703. The memory 702 stores machine-readable instructions that can be executed by the processor 701. When the computer device 700 is running, the processor 701 communicates with the memory 702 through the bus 703. When the machine-readable instructions are executed by the processor 701, the steps of the cross-frame jump trajectory recovery method shown in Embodiment 1 are performed.

[0133] Example 4 Based on the same concept, this application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the cross-frame jump trajectory recovery method described in any of the above embodiments.

[0134] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0135] The computer program product for cross-frame jump trajectory recovery provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0136] The cross-frame jump trajectory recovery device provided in this application embodiment can be specific hardware on a device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0137] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0138] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0139] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0140] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0141] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0142] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for recovering cross-frame jump trajectories, characterized in that, include: Obtain the target detection results in the current video frame; Obtain historical trajectory information of each target that has appeared in the video; Based on the historical trajectory information, anomaly analysis is performed on the motion state of each target in the current frame to determine whether cross-frame jump motion has occurred. When a target is determined to have a cross-frame jump motion, a cascaded trajectory recovery mechanism is triggered to generate the predicted trajectory of each target. The cascaded trajectory recovery mechanism includes at least two recovery stages executed in a preset order, wherein the later stage is triggered when the recovery of the previous stage fails. If no cross-frame jump motion is determined, the predicted trajectories of the targets are generated based on the conventional motion prediction model; The target detection results are matched with the predicted trajectories of each target to determine the target identity corresponding to each target detection result; Based on the determined target identity, the historical trajectory information of the corresponding target is updated using the target detection results to maintain continuous tracking of the target.

2. The method according to claim 1, characterized in that, The determination of whether a cross-frame jump motion has occurred includes determining whether a cross-frame jump motion has occurred in response to one or more of the following conditions being met: The target's displacement suddenly increases beyond a preset displacement threshold in N consecutive frames; The Intersection over Union (IoU) between the target and the detection results remains zero in the current frame and in consecutive preceding frames; The target maintains a consistent direction of motion across multiple consecutive frames and its rate of velocity change is below a preset rate of change threshold. The change in the target's detection bounding box size is less than a preset size threshold.

3. The method according to claim 1, characterized in that, The cascaded trajectory recovery mechanism includes: The first stage involves predicting and recovering the trajectory based on motion information from historical trajectories, generating a preliminary predicted trajectory, and determining the initial position of the target in the current frame based on the preliminary predicted trajectory. If the preliminary predicted trajectory cannot establish a valid match with any target detection result, the first stage recovery is deemed to have failed. Second stage: When the recovery of the first stage fails, a local image patch multi-scale search is performed in the vicinity of the initial position, and the initial position is corrected according to the search matching results to obtain the final predicted trajectory.

4. The method according to claim 3, characterized in that, The trajectory prediction and recovery based on historical trajectory motion information includes: Extract the target's location sequence from the historical trajectory information; A motion estimation model is constructed based on the position sequence to calculate the target's motion speed and motion trend; Based on the motion speed and motion trend, the initial position of the target in the current frame is predicted using interpolation or filtering methods.

5. The method according to claim 3, characterized in that, The local image patch multi-scale search includes: Extract the target's appearance features from historical trajectory information as a matching template; Construct a multi-scale image pyramid for the current frame; On each scale layer of the image pyramid, candidate search regions are delineated based on the initial position; Within the candidate search area, the similarity between each candidate window and the matching template is calculated using a sliding window method; The location of the candidate window with the highest similarity is selected as the corrected target location.

6. The method according to claim 3, characterized in that, Also includes: The final predicted trajectory is verified for consistency of motion direction and trajectory curvature constraint. If the verification passes, it is output as the predicted trajectory of each target. If the verification fails, the predicted trajectory is discarded and marked as recovery failure. It also includes trajectory lifecycle management: For target detection results that do not match any existing trajectory, initialize new historical trajectory information for them; for existing trajectories that do not match any detection results for multiple consecutive frames, remove them from the tracking queue.

7. The method according to claim 1, characterized in that, The step of matching the target detection results with the predicted trajectories of each target includes: Calculate the intersection-union ratio, center point distance, and appearance similarity between each detection box and each predicted trajectory; The intersection-union ratio, the center point distance, and the appearance similarity are fused to obtain a comprehensive matching score; Based on the comprehensive matching score, a one-to-one correspondence between the detection results and the predicted trajectory is determined by a bipartite graph matching algorithm, and an identity is assigned to the target.

8. A cross-frame jump trajectory recovery device, characterized in that, include: The detection and acquisition module is used to acquire the target detection results in the current video frame and the historical trajectory information of each target that has appeared in the video. The abnormal motion determination module is used to perform anomaly analysis on the motion state of each target in the current frame based on the historical trajectory information, and to determine whether cross-frame jump motion has occurred. The cascaded trajectory recovery module is used to trigger a cascaded trajectory recovery mechanism to generate predicted trajectories for each target when it is determined that the target has undergone cross-frame jump motion. The cascaded trajectory recovery mechanism includes at least two recovery stages executed in a preset order, wherein the latter stage is triggered when the former stage fails to recover. It is also used to generate predicted trajectories for each target based on a conventional motion prediction model if it is not determined that cross-frame jump motion has occurred. The matching and identity determination module is used to match the target detection results with the predicted trajectories of each target to determine the target identity corresponding to each target detection result; The trajectory update module is used to update the historical trajectory information of the corresponding target based on the determined target identity using the target detection results, so as to maintain continuous tracking of the target.

9. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the cross-frame jump trajectory recovery method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the cross-frame jump trajectory recovery method as described in any one of claims 1 to 7.

Images