Cloud-edge collaborative cross-camera target association method and system based on mobile logic chain

Abstract

The application provides a cloud-edge collaborative cross-camera target association method and system based on a mobile logic chain, which comprises the following steps: an edge device outputs structured trajectory metadata when a target enters or leaves a field of view; a cloud end constructs and outputs a mobile logic chain; the cloud end takes the disappearance event of the target reported by the edge device in real time and the mobile logic chain as input, calculates an adaptive time window according to the current real-time crowd density, determines that the mobile logic chain is broken when there is a strong logic edge and the appearance event of the target is not received within the adaptive time window, calculates a logic consistency score according to the cumulative number of trajectory breaks and a logic chain confidence factor reflecting the prediction accuracy of the edge; the cloud end outputs a scene state label through a scene state measurer, obtains a final trajectory credibility according to the edge observation credibility and the logic consistency, and performs an operation on the unique identifier of the target according to the relationship between the final trajectory credibility and a preset threshold. The application can keep the clarity of the cloud-edge responsibility boundary.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a cloud-edge collaborative cross-camera target association method and system based on mobile logic chain. Background Technology

[0002] The core task of target tracking is to maintain a unique identification number (ID) for each target across consecutive video frames. In single-camera scenarios, mainstream methods (such as DeepSORT and ByteTrack) achieve short-term ID stability by combining bounding boxes with motion models (such as Kalman filtering). However, once the target leaves the field of view or enters an occluded area, the trajectory terminates, making continuous tracking across cameras a core challenge in the field.

[0003] To achieve continuous tracking across cameras, academia and industry have conducted long-term research, resulting in two main technical approaches: The first approach: Feature matching based on Re-Identification (ReID). This method extracts the appearance features of a target (such as clothing, body shape, and other visual semantic information) at the edge or in the cloud, and matches the occurrence of the same target from different viewpoints by calculating cross-camera feature similarity. ReID methods offer excellent accuracy and are currently the mainstream technology for cross-camera tracking. However, as gait, body shape, and other personally identifiable information are explicitly classified as sensitive biometric information, their cross-device transmission requires individual consent. In scenarios such as public security and campus surveillance where individual consent cannot be obtained, the cross-domain transmission of ReID features faces compliance challenges. Maintaining cross-camera tracking accuracy while meeting privacy compliance has become a real constraint that this approach must address.

[0004] The second approach: a pure spatiotemporal geometric matching method using cloud-edge collaboration. To mitigate privacy compliance risks and reduce bandwidth consumption, a cloud-edge collaborative architecture has emerged in recent years: the edge only uploads lightweight detection results (such as bounding boxes, velocity, and location), while the cross-camera association task is handled by the cloud. Typical solutions (such as EdgeMOT) rely on spatiotemporal geometric constraints (such as transfer time, motion direction, and velocity consistency) for matching. This type of method has a natural advantage in meeting compliance requirements, but it has a fundamental limitation: it assumes that the target moves freely and the path has no prior knowledge, making it unable to distinguish between "missing reasonable paths" and "loss of the real ID." For example, in a campus scenario, "teaching building → cafeteria" is a historically frequent transfer path, while "teaching building → computer room" transfers are extremely rare; pure spatiotemporal models lack the ability to model such differences in transfer frequency, resulting in a high ID jump rate in complex topologies.

[0005] Common bottlenecks of existing methods: In recent years, researchers have attempted to introduce scene structure priors into cross-camera tracking (such as research related to CVPR and ICCV), achieving some progress by constructing scene topology graphs to guide associations. However, existing cloud-edge collaborative schemes generally adopt linear weighted models when fusing edge observation confidence with spatiotemporal constraints (or logical priors), assuming that their contributions are independent and their weights are fixed. This design exposes deep flaws in practical applications: state coupling effect: Under low-density steady-state conditions, the moving logic chain is highly reliable and should be trusted first; under high-density congestion or chaotic motion conditions, the predictive ability of the logic chain itself decreases, and the observation quality deteriorates accordingly, exhibiting a nonlinear coupling relationship. The linear model cannot dynamically adjust the decision logic according to the scene state, making it difficult to balance the ID jump rate and false retention rate in extreme scenarios.

[0006] In summary, overcoming the fundamental flaw of existing methods—that linear fusion of observation and logic cannot adapt to changes in scene state—while meeting privacy compliance constraints, and achieving high-precision cross-camera target association remains a critical technical challenge that urgently needs to be overcome in the field of intelligent video analytics. Summary of the Invention

[0007] Therefore, the purpose of this invention is to provide a cloud-edge collaborative cross-camera target association method and system based on mobile logic chain, so as to at least solve the shortcomings of the above-mentioned technology.

[0008] This invention proposes a cloud-edge collaborative cross-camera target association method based on mobile logical chains, comprising: Edge devices take real-time video streams as input, and through target detection and short-term tracking, generate and output structured trajectory metadata when a target enters or leaves the camera's field of view, and report it to the cloud. The structured trajectory metadata includes edge observation confidence, event type, timestamp, location information, and movement parameters. The cloud platform takes historical structured trajectory metadata as input, and constructs and outputs a mobile logical chain through statistical learning and spatiotemporal comparative encoding. In the mobile logic chain In the middle, node Represents region-event pairs, with directed edges. Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of the transfer path is calculated by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. The cloud uses the target data reported in real time by the edge device at the node. The disappearance event and the movement logic chain As input, an adaptive time window is calculated based on the current real-time pedestrian density, when a strong logical edge exists. And no target was received at the node within the adaptive time window. When an event occurs, it is determined that a movement logic chain break has occurred. The logic consistency score is calculated and output by taking the cumulative number of trajectory breaks and the logic chain confidence factor that reflects the edge prediction accuracy as inputs. The cloud takes the current area's pedestrian density, speed standard deviation, and trajectory break frequency, which are statistically analyzed in real time from the structured trajectory metadata, as input. It outputs scene status labels through a scene status calculator. Using the scene status labels as decision control signals, it performs mode switching fusion on the edge observation confidence and the logical consistency score, calculates and outputs the final trajectory credibility, and performs continuation, cache waiting, or termination operations on the unique identifier of the target based on the relationship between the final trajectory credibility and a preset threshold.

[0009] Furthermore, the formula for calculating the edge weight is as follows:

[0010] In the formula, Indicates the frequency of historical transitions; For nodes With embedding vector The cosine similarity, where, The node embedding vectors are generated by a lightweight spatiotemporal encoder to form structured attributes based on location, time period, and pedestrian density. For nodes and nodes Statistical dependence; , , For preset weights, and , .

[0011] Furthermore, the formula for calculating the adaptive time window is as follows:

[0012] In the formula, Represents the historical structured trajectory metadata from nodes To the node The average transfer time, Represents the historical structured trajectory metadata from nodes To the node The standard deviation of transfer time; As a dynamic adjustment factor, The current population density in the area. This represents the historical average population density for the current area.

[0013] Furthermore, the formula for calculating the logical consistency score is as follows:

[0014] In the formula, For the logical chain confidence factor, from the nodes in the past 24 hours The disappearing target is finally at the node The proportion of occurrence, The attenuation coefficient is... This indicates the cumulative number of trajectory breaks for the target.

[0015] Furthermore, the scene status labels are divided in the following manner: when , , When this occurs, the scene state label is output as steady state; when , , When this occurs, the scene state label is output as crowded. when , , When this happens, the scene state label is output as chaotic. in, The current population density in the area. For the speed standard deviation, The frequency of trajectory breakage. , , The steady-state threshold, , The threshold for disordered states.

[0016] Furthermore, the formula for calculating the reliability of the final trajectory is as follows: When the scenario state label is in a steady state, the logical consistency score is used. and the edge observation confidence As input, output the final trajectory reliability. ; When the scenario state label is "crowded," the logical consistency score is used. and the observation confidence adjusted by the time stretching function As input, output the final trajectory confidence level. ; When the scene state label is chaotic, the edge observation confidence level is used. and geographic similarity function As input, output the final trajectory confidence level. ; In the formula, For the fusion weights under steady state, Here, the fusion weights are for the crowded state, where, , ; , This is the adjustment coefficient; , Indicates the last known location of the target. Indicates the next location where the target will appear. Indicates spatial bandwidth parameter, This is the attenuation factor.

[0017] Furthermore, based on the relationship between the final trajectory reliability and a preset threshold, the steps of performing continuation, caching, or termination operations on the unique identifier of the target include: When the credibility of the final trajectory When this occurs, output the target's unique identifier continuation command; When the credibility of the final trajectory When the target's unique identifier is retained, the waiting time is extended, and a query command is sent to the edge side to request the retrospective of recent low-quality detection results. The credibility of the final trajectory When the time comes, output the target's unique identifier termination command.

[0018] This invention also proposes a cloud-edge collaborative cross-camera target association system based on a mobile logical chain, comprising: The edge tracking module is used to control edge devices to take real-time video stream as input, and through target detection and short-term tracking, generate and output structured trajectory metadata when the target enters or leaves the camera's field of view, and report it to the cloud. The structured trajectory metadata includes edge observation confidence, event type, timestamp, location information and movement parameters. The mobile logic chain construction module controls the cloud to construct and output a mobile logic chain by taking historical structured trajectory metadata as input and through statistical learning and spatiotemporal comparative encoding. In the mobile logic chain In the middle, node Represents region-event pairs, with directed edges. Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of the transfer path is calculated by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. The logic chain break detection module is used to control the target reported in real time by the cloud and the edge device at the node. The disappearance event and the movement logic chain As input, an adaptive time window is calculated based on the current real-time pedestrian density, when a strong logical edge exists. And no target was received at the node within the adaptive time window. When an event occurs, it is determined that a movement logic chain break has occurred. The logic consistency score is calculated and output by taking the cumulative number of trajectory breaks and the logic chain confidence factor that reflects the edge prediction accuracy as inputs. The scene state measurement and decision-making module is used to control the cloud to take the current area's pedestrian density, speed standard deviation, and trajectory break frequency, which are statistically analyzed in real time from the structured trajectory metadata, as input, and output scene state labels through the scene state measurer; using the scene state labels as decision control signals, it performs mode switching fusion on the edge observation confidence and the logical consistency score, calculates and outputs the final trajectory credibility, and performs continuation, cache waiting, or termination operations on the unique identifier of the target based on the relationship between the final trajectory credibility and a preset threshold.

[0019] The present invention also proposes a storage medium on which a computer program is stored, which, when executed by a processor, implements the above-described cloud-edge collaborative cross-camera target association method based on mobile logic chain.

[0020] The present invention also proposes a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described cloud-edge collaborative cross-camera target association method based on mobile logic chain.

[0021] Compared with existing technologies, the cloud-edge collaborative cross-camera target association method and system based on mobile logical chain in this invention has the following advantages: (1) Meets compliance requirements: The edge device only uploads structured trajectory metadata that conforms to the standard, and does not transmit original video, image cropping or biometric vectors, which complies with regulations and is compatible with video networking standards; (2) Improve ID continuity: In actual test environment, after adopting this method, the cross-camera target ID jump rate is significantly reduced, and the corresponding ID continuity maintenance rate is improved, which meets the indicator requirement of cross-domain tracking success rate of not less than 90% in communication standard; especially in extreme scenarios such as high-density congestion and chaotic motion, through the adaptive decision-making mechanism of scene state calculation, the ID jump rate is significantly reduced compared with the fixed weight linear fusion model, which verifies the robustness of this method in complex scenarios; (3) Reduce network bandwidth usage: The average reported data volume of a single target event is less than 1 KB. Under typical deployment conditions, it supports a single cloud node to handle concurrent reporting of tens of thousands of edge devices. (4) Traceability: Each ID state change is associated with a specific logical chain break event and the corresponding adaptive time window timeout record, and records the scene state label when the decision is triggered: steady state, crowded state, chaotic state, and the mode switching branch adopted, which facilitates post-event verification and parameter tuning; (5) Cloud and edge functions are interdependent: Edge devices are limited by memory capacity (not exceeding 64 MB) and cannot run graph neural networks or high-dimensional feature extraction models. They need to rely on the mobile logic chain provided by the cloud for cross-regional inference. The cloud does not access the original video stream and cannot perform local target tracking. It needs to rely on the structured trajectory metadata generated by the edge. The statistical quantities required by the scene state estimator of this method are all calculated in real time by the cloud based on the metadata reported by the edge. No additional edge capabilities are required, which maintains the clarity of the cloud and edge responsibility boundaries. Attached Figure Description

[0022] Figure 1 This is a flowchart of the cloud-edge collaborative cross-camera target association method based on mobile logic chain in the first embodiment of the present invention; Figure 2 This is a comparison chart of ID jump rate and false retention rate between the cloud-edge collaborative cross-camera target association method based on mobile logic chain and EdgeMOT under different crowd densities in the first embodiment of the present invention. Figure 3 This is a structural block diagram of the cloud-edge collaborative cross-camera target association system based on mobile logic chain in the second embodiment of the present invention; Figure 4 This is a structural block diagram of the computer in the third embodiment of the present invention.

[0023] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0024] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0026] Example 1 Please see Figure 1 The figure shows a cloud-edge collaborative cross-camera target association method based on a mobile logical chain in the first embodiment of the present invention. The method specifically includes steps S101 to S104: S101, the edge device takes real-time video stream as input, and through target detection and short-term tracking, generates and outputs structured trajectory metadata and reports it to the cloud when the target enters or leaves the camera's field of view. The structured trajectory metadata includes edge observation confidence, event type, timestamp, location information and movement parameters. In practical implementation, this method is applied to controlled security scenarios such as campuses, airports, and water conservancy projects. The method includes edge devices (e.g., camera equipment) located at the network edge and a service cluster deployed in the cloud. At the edge, each camera is responsible for capturing video streams and performing preliminary processing on the video content through the edge devices. Structured trajectory metadata conforming to video image information application standards is extracted and uploaded. The original video, image cropping, or biometric vectors are removed from the domain. The structured trajectory metadata only contains standardized event records of the target's spatiotemporal behavioral semantics (such as location, time, speed, and direction), and does not contain any pixel-level visual content or biometric features that can identify an individual. This metadata serves as the sole observation input for cloud-based mobile logic chain inference, containing key semantics of the target's spatiotemporal behavior but without carrying any visual information that can identify an individual.

[0027] On each edge computing node (such as an industrial-grade AI server that supports 20 channels of 1080p video access), a lightweight object detection and short-term tracking component (such as a simplified version of the video object tracker ByteTrack) is run. When a target enters or leaves the field of view, it generates a JSON (JavaScript Object Notation) format event that conforms to the video image information application standard.

[0028] The metadata is encapsulated in standardized JSON (JavaScript Object Notation) format and includes the following fields: target_local_id: Target local ID (string), a temporary unique identifier assigned by the edge device to the current target; event_type: Event type, with values ​​of "appear" or "disappear", indicating whether the target enters or leaves the current camera's field of view, respectively; timestamp: The timestamp of the event (Unix milliseconds), indicating the moment the event occurred (UTC or local system time); geohash: GeoHash encoded location: Indicates the geocoded location of the target when the event occurred, with an accuracy of ≤ 50 meters; speed: Instantaneous velocity (unit: m / s), estimated by ByteTrack deployed on the edge device based on consecutive frames; direction: Direction of movement: Starting from true north as 0°, increase clockwise; confidence: confidence level of edge observations (Floating-point number, ranging from 0 to 1), calculated by the edge target tracker based on factors such as detection box score, number of consecutive trajectory frames, and motion consistency, used to quantify the reliability of edge detection and short-term tracking results. A higher value indicates more reliable edge observation, and the cloud will use this as the basic confidence input when associating across cameras.

[0029] The JSON object, after compression, has an average size of less than 1 KB and can be securely uploaded to the cloud via a protocol channel or HTTPS interface. The original video frames, image cropping, and feature vectors are not uploaded.

[0030] S102, the cloud takes historical structured trajectory metadata as input, and constructs and outputs a mobile logical chain through statistical learning and spatiotemporal comparison encoding. In the mobile logic chain In the middle, node Represents region-event pairs, with directed edges. Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of the transfer path is calculated by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. In this embodiment, the movement logic chain refers to a computable prior model that represents the high-frequency spatial movement patterns of targets in a controlled scene, obtained through statistical learning from historical normal traffic data. Its core function is to provide a structured semantic basis for cross-camera target association without relying on any visual features of the target's appearance.

[0031] In practice, load balancers (such as Nginx or application-specific load balancers) distribute metadata requests from edge devices to the most idle servers within the cloud cluster, ensuring stable and orderly processing of massive tasks under tens of thousands of concurrent connections. A mobile logical chain is then built on the cloud application server cluster based on historical trajectory metadata. In the middle, node This represents a region-event pair, such as "Someone left the east gate of the teaching building"; Directed edge Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of a transfer path is quantified by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. The edge weights, calculated by combining these three factors, reflect the reliability of the transfer. To enhance semantic representation, a lightweight spatiotemporal contrastive learning encoder is designed in this embodiment. (Constituted by a two-layer fully connected neural network with fewer than 10KB of parameters), mapping each region event to a low-dimensional embedding vector:

[0032] in, For nodes Embedded vector, For the encoding function, the input is structured attributes (location, time of day, current area population density).

[0033] The formula for calculating the edge weight is as follows:

[0034] In the formula, Indicates the historical transfer frequency (from node in the past 30 days). After disappearing, in the embedding vector (proportion of occurrence), the larger the value, the more frequently it occurs; For nodes With embedding vector Cosine similarity (a measure of semantic similarity), where, The node embedding vectors are generated by a lightweight spatiotemporal encoder to form structured attributes based on location, time period, and pedestrian density. For nodes and nodes Statistical dependence; , , For preset weights, and , This is to ensure that semantic information is incorporated. In this embodiment, , , Choose 0.5, 0.3, 0.2.

[0035] S103, the cloud uses the target reported in real time by the edge device at the node. The disappearance event and the movement logic chain As input, an adaptive time window is calculated based on the current real-time pedestrian density, when a strong logical edge exists. And no target was received at the node within the adaptive time window. When an event occurs, it is determined that a movement logic chain break has occurred. The logic consistency score is calculated and output by taking the cumulative number of trajectory breaks and the logic chain confidence factor that reflects the edge prediction accuracy as inputs. In this embodiment, the mobile logic chain break event is a quantifiable and traceable decision trigger signal: when the target disappears along the historical high-frequency transfer path and does not reappear in the expected next area within the adaptive time window, it indicates that the system prediction has failed (rather than the physical loss of the target). This event is directly used to drive the correction or termination of the ID state.

[0036] In practical implementation, when the edge reports "target disappears at node", "After that, in the adaptive time window" No "target appeared at node" received The report was submitted, and there was a strong logical boundary. (Right now If the value exceeds a preset threshold, it is determined that a movement logic chain has been broken.

[0037] Furthermore, the formula for calculating the adaptive time window is as follows:

[0038] In the formula, Represents the historical structured trajectory metadata from nodes To the node The average transfer time, Represents the historical structured trajectory metadata from nodes To the node The standard deviation of transfer time; As a dynamic adjustment factor, The current population density in the area. This represents the historical average population density for the current area.

[0039] Furthermore, considering the differences in the historical prediction reliability of migration paths in different regions, this embodiment introduces a logistic chain confidence factor. The logical consistency score is calibrated, specifically the logical chain confidence factor. Defined as within a preset time window (in this embodiment, the window is 24 hours), from node Among the missing targets, success lies at the node. The proportion of occurrence is used to reflect the edges in the movement logic chain. The actual prediction accuracy is recorded; the lower the value, the less reliable the transfer path. The number of trajectory breaks is also recorded. And calculate the calibrated logical consistency score:

[0040] In the formula, For the logical chain confidence factor, from the nodes in the past 24 hours The disappearing target is finally at the node The proportion of occurrence, The attenuation coefficient controls the severity of the penalty imposed on the credibility of the fracture event. The cumulative number of trajectory breaks of the target, and the logical consistency score. The ID state decreases smoothly with increasing number of fractures, avoiding abrupt changes in the ID state.

[0041] Specifically, if the target transfer path has no corresponding high-frequency transfer record in the historical traffic data (i.e., there is no valid logical edge), If the ID is not supported by the mobile logic chain, it will be directly judged as an unreliable association and the ID will be actively terminated to avoid accidental merging across cameras.

[0042] S104, the cloud takes the current area's pedestrian density, speed standard deviation, and trajectory break frequency, which are statistically analyzed in real time from the structured trajectory metadata, as input, and outputs a scene status label through a scene status calculator; using the scene status label as a decision control signal, it performs mode switching fusion on the edge observation confidence and the logical consistency score, calculates and outputs the final trajectory credibility, and performs continuation, cache waiting, or termination operations on the unique identifier of the target based on the relationship between the final trajectory credibility and a preset threshold.

[0043] In practical implementation, a scene state calculator is introduced. This sensor takes the real-time pedestrian density, speed standard deviation, and trajectory break frequency of the current area as input, and outputs the current scene state label to achieve adaptive matching between decision logic and scene dynamics. All three parameters are calculated in real-time by the cloud based on metadata reported from the edge. The current population density in the area is calculated from cloud-based statistics on target events that have not yet disappeared within that area. The standard deviation of the instantaneous velocity of all targets within the region is calculated by the cloud based on the velocity values ​​reported by each target. The average trajectory breakage frequency of the edge tracker over the past 30 seconds is derived from the proportion of events that did not appear in the expected area among the disappearance events statistically analyzed in the cloud.

[0044] Furthermore, the scene status labels are divided in the following manner: when , , When the scene state label is output as steady state, the crowd is sparse, the movement is orderly, the observation quality is high, the movement logic chain is highly reliable, and the decision should prioritize trusting the logic consistency score. when , , When the scene is crowded, the output scene state label is crowded. At this time, the crowd density is high but the movement is orderly (such as a queuing scene). The movement logic chain is still effective, but the observation quality is reduced due to occlusion. The decision should maintain trust in the logic chain, but appropriately extend the waiting window to adapt to the delay caused by congestion. when , , When the scene state label is output as chaotic, high density and high chaos coexist. The predictive ability of the mobile logic chain itself decreases and the observation quality is extremely low. The decision should turn to trust "spatial proximity" rather than forcibly associate IDs. If necessary, actively downgrade and request manual intervention. in, The current population density in the area. For the speed standard deviation, The frequency of trajectory breakage. , , The steady-state threshold, , In this embodiment, the threshold for disordered states is used. , , The values ​​were 15, 0.8, and 0.2 respectively. , The values ​​are 30, 1.5, and 0.5 respectively.

[0045] Based on the state measurement results, this embodiment adopts a mode switching decision mechanism, and the formula for calculating the reliability of the final trajectory is as follows: When the scenario state label is in a steady state, the logical consistency score is used. and the edge observation confidence As input, output the final trajectory confidence level. ; When the scenario state label is "crowded," the logical consistency score is used. and the observation confidence adjusted by the time stretching function As input, output the final trajectory confidence level. ; When the scene state label is chaotic, the edge observation confidence level is used. and geographic similarity function As input, output the final trajectory confidence level. ; In the formula, For the fusion weights under steady state, Here, the fusion weights are for the crowded state, where, , ; As a time-spanning function, it dynamically adjusts the contribution of observation confidence based on crowd density under congested conditions. The adjustment coefficient is preferably 0.3 in this embodiment; , Indicates the last known location of the target. Indicates the next location where the target will appear. This represents the spatial bandwidth parameter; in this embodiment, the value is selected as 50. The attenuation factor is set to 0.8.

[0046] In a chaotic state, the minimum value between the observation confidence score and the geographical similarity decay value is taken as the final trajectory confidence score, achieving a conservative decision-making approach that prioritizes terminating the ID rather than perpetuating errors. Specifically, based on the relationship between the final trajectory confidence score and a preset threshold, the steps for performing continuation, caching, or termination operations on the unique identifier of the target include: When the credibility of the final trajectory When this occurs, output the target's unique identifier continuation command; When the credibility of the final trajectory When the target's unique identifier is retained, the waiting time is extended, and a query command is sent to the edge side to request the retrospective of recent low-quality detection results. The credibility of the final trajectory When the time comes, output the target's unique identifier termination command.

[0047] Furthermore, to evaluate the technical effectiveness of this embodiment, the following three representative prior art technologies are selected as baselines: (1) Single-camera target detection and short-time tracking method (represented by ByteTrack) ByteTrack is a widely deployed real-time target tracking algorithm. Its core idea is to predict the target position using Kalman filtering and maintain the ID by utilizing high-resolution matching between the detection box and the trajectory. This method relies on only a single video stream and does not perform cross-camera correlation. Once the target leaves the field of view for more than a preset time (usually 2 seconds), the trajectory is terminated. In controlled scenes with blind spots or occlusions, this method cannot handle cross-regional continuity issues, has a high ID jump rate, and struggles to meet tracking rate requirements.

[0048] (2) Traditional cloud-edge collaborative tracking method (represented by EdgeMOT) EdgeMOT-like solutions deploy lightweight detection at the edge and move trajectory association tasks to the cloud to reduce bandwidth. Their cross-camera matching primarily relies on spatiotemporal geometric constraints (such as velocity, direction, and transition time windows). However, these methods assume free and irregular target movement and do not introduce prior scene traversal logic. When multiple legitimate paths coexist or abnormal transitions occur, the system cannot distinguish between reasonable missing paths and genuine loss, leading to insufficient ID stability in complex topologies.

[0049] (3) High-bandwidth ReID feature upload scheme This solution extracts re-identification (ReID) feature vectors (typically 256–512 dimensions) of targets at the edge and uploads them to the cloud for cross-camera similarity matching. While the ID accuracy is high, the bandwidth consumption per channel reaches tens of KB / s, making it unscalable in scenarios with thousands of concurrent channels. More importantly, ReID features are biometric information, and their cross-device transmission does not comply with regulations, making compliant deployment feasible.

[0050] The method in this embodiment constructs a dedicated dataset based on structured trajectory metadata collected from a real university security system. The data covers 30 days and 20 entrance / exit cameras, including fields such as target ID, appearance / disappearance timestamp, GeoHash location, speed, and direction, strictly adhering to standard formats. To provide reliable tags, access control card swipe records are synchronously integrated, using the "card swipe to leave campus" event as the criterion for determining the target's actual departure, and is used to calculate the false retention rate. All trajectories are manually sampled and verified to ensure spatiotemporal consistency.

[0051] It should be noted that due to campus security and privacy concerns, the original videos have not been publicly released, but the structured metadata has been anonymized and can be used for internal verification. This method is not dependent on any specific scenario and can be transferred to similar controlled environments such as construction sites, hydroelectric power stations, and subway stations. The technical parameters of the self-built dataset are listed in Tables 1 and 2.

[0052] Table 1. Key Statistical Information of the Self-Built Campus Scene Multi-Source Trajectory Dataset

[0053] Table 2 Detailed technical parameters of self-built dataset

[0054] Taking student ID=1001 leaving from the east gate of teaching building A as an example: 1. When the edge device detects the disappearance of the target, it generates a standard-compliant metadata event and reports it. 2. Query the mobile logical chain in the cloud, identify the strong transition edge "Teaching Building A → North Section of Main Road", and obtain the parameters: edge weight. Logical chain confidence factor ,from arrive average transfer time Corresponding standard deviation ; 3. Calculate the adaptive time window based on the current pedestrian density. And if no event is received in the window, the logic chain is determined to be broken; 4. Calculate the logical consistency score The final trajectory confidence level is obtained by fusing the observation confidence levels. ; 5. Because Keep the ID and continue waiting.

[0055] To demonstrate the advancement of this embodiment over the prior art, it is compared end-to-end with the three representative baseline methods mentioned above on the same campus scenario dataset (Table 3).

[0056] Table 3 Performance comparison between the method in this embodiment and a representative prior art baseline

[0057] All methods were evaluated on the same edge device (memory ≤ 64MB) and campus scene dataset. ByteTrack could not achieve cross-camera tracking, so its "ID jump rate" statistic is the proportion of trajectory breakage in cross-region tasks (i.e., the proportion of the system that cannot recover the same ID in the next camera after the target leaves the current camera's field of view). The value reflects the breakage situation in actual tests. Table 3 shows that the method in this embodiment significantly outperforms existing technologies while fully meeting privacy compliance and low bandwidth constraints. Compared with the current mainstream cloud-edge collaboration solution EdgeMOT, the ID hopping rate is reduced from 22.0% to 10.8%, a decrease of 51%. Furthermore, the method in this embodiment is the only one that effectively controls the risk of "excessive continuation" (false retention rate of 3.6%) within a compliant solution. ByteTrack lacks cross-domain capabilities and cannot meet practical needs; while the ReID solution offers high accuracy, uploading biometric features violates regulations and is not feasible for deployment.

[0058] Furthermore, to verify the contribution of each technical component in this embodiment, an ablation experiment was designed, and the results are shown in Table 4: Table 4. Impact of each component of the method in this embodiment on ID stability and false retention rate (ablation experiment results)

[0059] Among them, "no mobile logical chain" refers to the technical solution that does not adopt the above-mentioned mobile logical chain mechanism and only relies on spatiotemporal geometric constraints for cross-camera association; "upload ReID features" corresponds to the high-bandwidth re-identification solution. Although the ID accuracy is high, it is not feasible for practical deployment because uploading the original features violates the requirement that "the original video and identifiable information do not leave the domain". The false retention rate is defined as the proportion of targets that have actually left the system (e.g., left the school), but the system still incorrectly maintains their IDs without terminating them, reflecting the risk of "over-continuation" of the trajectory; All solutions were tested on the same edge device (≤64MB memory) and dataset to ensure fair comparison.

[0060] Experimental results show that the method in this embodiment achieves an optimal balance between ID stability and decision robustness. Despite "no "The proposed solution has a slightly lower ID jump rate (10.5% vs 10.8%), but its false retention rate is as high as 5.8%, indicating that the lack of confidence calibration leads to the system over-trusting unreliable paths; a logical chain confidence factor is introduced." The false retention rate was significantly reduced to 3.6%. However, removing the mobile logical chain or using a fixed time window resulted in an ID jump rate exceeding 19%, failing to meet the standard requirements. In summary, mobile logical chain modeling, adaptive time windows, and logical chain confidence factors are crucial. These three elements constitute an inseparable technical whole, jointly supporting the excellent performance of the method in this embodiment under compliance constraints.

[0061] Furthermore, to verify the robustness of this embodiment under different environmental pressures, a dynamic experiment was designed, using real-time pedestrian density within a single field of view as the variable, to compare the performance of the method in this embodiment with EdgeMOT in terms of ID jump rate and false retention rate. Wherein: pedestrian density = number of targets detected within the current field of view of a single camera (unit: people / field of view, 30 people / field of view is considered severely crowded); ID jump rate reflects the continuity of IDs across cameras, and a lower value is better; false retention rate (%) reflects the error of "the target has left but the system still maintains the ID," and a lower value is better.

[0062] Experimental structure as follows Figure 2 As shown in the figure, pedestrian density is the independent variable (X-axis), and ID jump rate and false retention rate are the dependent variables (located on the left Y-axis and right Y-axis, respectively).

[0063] In this embodiment: as the pedestrian density increases, the ID hopping rate slowly rises from 8.2% to 10.8%, and the false retention rate rises from 2.4% to 3.6%. EdgeMOT: The ID hopping rate deteriorates from 15.6% to 22.8%, while the false retention rate remains stable at 1.8%~2.4%. Under high density (≥20 people / field of view), the ID hopping rate of this embodiment is 10-12 percentage points lower than that of EdgeMOT, demonstrating the robustness of the mobile logic chain in complex scenarios.

[0064] In summary, the cloud-edge collaborative cross-camera target association method based on mobile logical chain in the above embodiments of the present invention has the following beneficial effects: (1) Meets compliance requirements: The edge device only uploads structured trajectory metadata that conforms to the standard, and does not transmit original video, image cropping or biometric vectors, which complies with regulations and is compatible with video networking standards; (2) Improve ID continuity: In actual test environment, after adopting this method, the cross-camera target ID jump rate is significantly reduced, and the corresponding ID continuity maintenance rate is improved, which meets the indicator requirement of cross-domain tracking success rate of not less than 90% in communication standard; especially in extreme scenarios such as high-density congestion and chaotic motion, through the adaptive decision-making mechanism of scene state calculation, the ID jump rate is significantly reduced compared with the fixed weight linear fusion model, which verifies the robustness of this method in complex scenarios; (3) Reduce network bandwidth usage: The average reported data volume of a single target event is less than 1 KB. Under typical deployment conditions, it supports a single cloud node to handle concurrent reporting of tens of thousands of edge devices. (4) Traceability: Each ID state change is associated with a specific logical chain break event and the corresponding adaptive time window timeout record, and records the scene state label when the decision is triggered: steady state, crowded state, chaotic state, and the mode switching branch adopted, which facilitates post-event verification and parameter tuning; (5) Cloud and edge functions are interdependent: Edge devices are limited by memory capacity (not exceeding 64 MB) and cannot run graph neural networks or high-dimensional feature extraction models. They need to rely on the mobile logic chain provided by the cloud for cross-regional inference. The cloud does not access the original video stream and cannot perform local target tracking. It needs to rely on the structured trajectory metadata generated by the edge. The statistical quantities required by the scene state estimator of this method are all calculated in real time by the cloud based on the metadata reported by the edge. No additional edge capabilities are required, which maintains the clarity of the cloud and edge responsibility boundaries.

[0065] Example 2 In another aspect, this invention proposes a cloud-edge collaborative cross-camera target association system based on a mobile logical chain. Please refer to [link / reference needed]. Figure 3The figure shows a cloud-edge collaborative cross-camera target association system based on a mobile logical chain according to a second embodiment of the present invention. The system includes: The edge tracking module 11 is used to control the edge device to take real-time video stream as input, and through target detection and short-term tracking, generate and output structured trajectory metadata and report it to the cloud when the target enters or leaves the camera's field of view. The structured trajectory metadata includes edge observation confidence, event type, timestamp, location information and movement parameters. The mobile logic chain construction module 12 is used to control the cloud to construct and output a mobile logic chain by taking historical structured trajectory metadata as input and through statistical learning and spatiotemporal comparison encoding. In the mobile logic chain In the middle, node Represents region-event pairs, with directed edges. Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of the transfer path is calculated by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. The logic chain break detection module 13 is used to control the target reported by the cloud and the edge device in real time at the node. The disappearance event and the movement logic chain As input, an adaptive time window is calculated based on the current real-time pedestrian density, when a strong logical edge exists. And no target was received at the node within the adaptive time window. When an event occurs, it is determined that a movement logic chain break has occurred. The logic consistency score is calculated and output by taking the cumulative number of trajectory breaks and the logic chain confidence factor that reflects the edge prediction accuracy as inputs. The scene state measurement and decision module 14 is used to control the cloud to take the current area's pedestrian density, speed standard deviation, and trajectory break frequency, which are statistically analyzed in real time from the structured trajectory metadata, as input, and output scene state labels through the scene state measurer; using the scene state labels as decision control signals, it performs mode switching fusion on the edge observation confidence and the logical consistency score, calculates and outputs the final trajectory credibility, and performs continuation, cache waiting, or termination operations on the unique identifier of the target based on the relationship between the final trajectory credibility and a preset threshold.

[0066] The functions or operation steps implemented by the above modules and units are largely the same as those in the above method embodiments, and will not be repeated here.

[0067] The cloud-edge collaborative cross-camera target association system based on mobile logic chain provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0068] Example 3 This invention also proposes a computer, please refer to [link / reference]. Figure 4 The computer shown in the third embodiment of the present invention includes a memory 10, a processor 20, and a computer program 30 stored in the memory 10 and executable on the processor 20. When the processor 20 executes the computer program 30, it implements the above-described cloud-edge collaborative cross-camera target association method based on mobile logic chain.

[0069] The memory 10 includes at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10 can be an internal storage unit of a computer, such as the computer's hard disk. In other embodiments, the memory 10 can be an external storage device, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Furthermore, the memory 10 can include both internal and external storage units of the computer. The memory 10 can be used not only to store application software and various types of data installed on the computer, but also to temporarily store data that has been output or will be output.

[0070] In some embodiments, the processor 20 may be an electronic control unit (ECU), a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip, used to run program code stored in the memory 10 or process data, such as executing access restriction programs.

[0071] It should be pointed out that, Figure 4 The structure shown does not constitute a limitation on the computer. In other embodiments, the computer may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0072] This invention also proposes a storage medium storing a computer program that, when executed by a processor, implements the cloud-edge collaborative cross-camera target association method based on a mobile logic chain as described above.

[0073] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0074] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0075] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A cloud-edge collaborative cross-camera target association method based on mobile logical chains, characterized in that, include: Edge devices take real-time video streams as input, and through target detection and short-term tracking, generate and output structured trajectory metadata when a target enters or leaves the camera's field of view, and report it to the cloud. The structured trajectory metadata includes edge observation confidence, event type, timestamp, location information, and movement parameters. The cloud platform takes historical structured trajectory metadata as input, and constructs and outputs a mobile logical chain through statistical learning and spatiotemporal comparative encoding. In the mobile logic chain In the middle, node Represents region-event pairs, with directed edges. Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of the transfer path is calculated by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. The cloud uses the target data reported in real time by the edge device at the node. The disappearance event and the movement logic chain As input, an adaptive time window is calculated based on the current real-time pedestrian density, when a strong logical edge exists. And no target was received at the node within the adaptive time window. When an event occurs, it is determined that a movement logic chain break has occurred. The logic consistency score is calculated and output by taking the cumulative number of trajectory breaks and the logic chain confidence factor that reflects the edge prediction accuracy as inputs. The cloud takes the current area's pedestrian density, speed standard deviation, and trajectory break frequency, which are statistically analyzed in real time from the structured trajectory metadata, as input. It outputs scene status labels through a scene status calculator. Using the scene status labels as decision control signals, it performs mode switching fusion on the edge observation confidence and the logical consistency score, calculates and outputs the final trajectory credibility, and performs continuation, cache waiting, or termination operations on the unique identifier of the target based on the relationship between the final trajectory credibility and a preset threshold.

2. The cloud-edge collaborative cross-camera target association method based on mobile logical chain according to claim 1, characterized in that, The formula for calculating the edge weight is: In the formula, Indicates the frequency of historical transitions; For nodes With embedding vector The cosine similarity, where, The node embedding vectors are generated by a lightweight spatiotemporal encoder to form structured attributes based on location, time period, and pedestrian density. For nodes and nodes Statistical dependence; , , For preset weights, and , .

3. The cloud-edge collaborative cross-camera target association method based on mobile logical chain according to claim 2, characterized in that, The formula for calculating the adaptive time window is: In the formula, Represents the historical structured trajectory metadata from nodes To the node The average transfer time, Represents the historical structured trajectory metadata from nodes To the node The standard deviation of transfer time; As a dynamic adjustment factor, The current population density in the area. This represents the historical average population density for the current area.

4. The cloud-edge collaborative cross-camera target association method based on mobile logical chain according to claim 3, characterized in that, The formula for calculating the logical consistency score is as follows: In the formula, For the logical chain confidence factor, from the nodes in the past 24 hours The disappearing target is finally at the node The proportion of occurrence, The attenuation coefficient is... This indicates the cumulative number of trajectory breaks for the target.

5. The cloud-edge collaborative cross-camera target association method based on mobile logical chain according to claim 4, characterized in that, The scene status labels are divided in the following way: when , , When this occurs, the scene state label is output as steady state; when , , When this occurs, the scene state label is output as crowded. when , , When this happens, the scene state label is output as chaotic. in, The current population density in the area. For the speed standard deviation, The frequency of trajectory breakage. , , The steady-state threshold, , The threshold for disordered states.

6. The cloud-edge collaborative cross-camera target association method based on mobile logical chain according to claim 5, characterized in that, The formula for calculating the reliability of the final trajectory is: When the scenario state label is in a steady state, the logical consistency score is used. and the edge observation confidence As input, output the final trajectory reliability. ; When the scenario state label is "crowded," the logical consistency score is used. and the observation confidence adjusted by the time stretching function As input, output the final trajectory reliability. ; When the scene state label is chaotic, the edge observation confidence level is used. and geographic similarity function As input, output the final trajectory reliability. ; In the formula, For the fusion weights under steady state, Here, the fusion weights are for the crowded state, where, , ; , This is the adjustment coefficient; , Indicates the last known location of the target. Indicates the next location where the target will appear. Indicates spatial bandwidth parameter, This is the attenuation factor.

7. The cloud-edge collaborative cross-camera target association method based on mobile logical chain according to claim 6, characterized in that, Based on the relationship between the final trajectory reliability and a preset threshold, the steps for performing continuation, caching, or termination operations on the unique identifier of the target include: When the credibility of the final trajectory When this occurs, output the target's unique identifier continuation command; When the credibility of the final trajectory When the time is reached, a unique identifier retention instruction for the target is output and the waiting time is extended. At the same time, a query instruction is sent to the edge side to request the retrospective of recent low-quality detection results. The credibility of the final trajectory When the time comes, output the target's unique identifier termination command.

8. A cloud-edge collaborative cross-camera target association system based on mobile logical chain, characterized in that, include: The edge tracking module is used to control edge devices to take real-time video stream as input, and through target detection and short-term tracking, generate and output structured trajectory metadata when the target enters or leaves the camera's field of view, and report it to the cloud. The structured trajectory metadata includes edge observation confidence, event type, timestamp, location information and movement parameters. The mobile logic chain construction module controls the cloud to construct and output a mobile logic chain by taking historical structured trajectory metadata as input and through statistical learning and spatiotemporal comparative encoding. In the mobile logic chain In the middle, node Represents region-event pairs, with directed edges. Indicates the target is from the region Transfer to the region Historical high-frequency traffic relationships, edge weights The reliability of the transfer path is calculated by integrating historical transfer frequency, cosine similarity of node embedding vectors, and normalized mutual information. The logic chain break detection module is used to control the target reported in real time by the cloud and the edge device at the node. The disappearance event and the movement logic chain As input, an adaptive time window is calculated based on the current real-time pedestrian density, when a strong logical edge exists. And no target was received at the node within the adaptive time window. When an event occurs, it is determined that a movement logic chain break has occurred. The logic consistency score is calculated and output by taking the cumulative number of trajectory breaks and the logic chain confidence factor that reflects the edge prediction accuracy as inputs. The scene state measurement and decision-making module is used to control the cloud to take the current area's pedestrian density, speed standard deviation, and trajectory break frequency, which are statistically analyzed in real time from the structured trajectory metadata, as input, and output scene state labels through the scene state measurer; using the scene state labels as decision control signals, it performs mode switching fusion on the edge observation confidence and the logical consistency score, calculates and outputs the final trajectory credibility, and performs continuation, cache waiting, or termination operations on the unique identifier of the target based on the relationship between the final trajectory credibility and a preset threshold.

9. A readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the cloud-edge collaborative cross-camera target association method based on mobile logic chain as described in any one of claims 1 to 7.

10. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the cloud-edge collaborative cross-camera target association method based on mobile logic chain as described in any one of claims 1 to 7.

Images