An open block cross-lens behavior chain reconstruction method and system

Abstract

The application provides an open block cross-lens behavior chain reconstruction method and system, relating to the technical field of computer vision, which constructs a functional space structure model in the target open block, generates a space unit reachable matrix and an adaptive transfer time window, establishes a monocular absolute scale mapping in combination with the camera parameters to restore the absolute ground physical trajectory of the target, thereby extracting absolute scale behavior features with dimensions in a unified physical coordinate system. By introducing a spatial boundary reference system, behavior preference anchor points are identified, and the weights of difference items in the associated cost function are dynamically adjusted according to the stability of historical preferences, and a pre-filtering mechanism is used to remove physically inaccessible candidate segments, finally realizing template updating, interruption recovery and cross-day chain maintenance of anonymous behavior chains. The technical scheme effectively reduces the risk of misassociation in open block cross-lens tracking, and significantly improves the stability, continuity and explainability of long-period behavior chain reconstruction.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a method and system for reconstructing cross-camera behavior chains in open blocks. Background Technology

[0002] In open street scenarios, due to the open space, discrete camera distribution, and repeated appearance of targets across cameras, existing technologies typically rely on appearance features, re-identification features, or coarse time windows for cross-camera association. However, these methods often struggle to effectively address issues such as occlusion, lighting changes, clothing variations, and disturbances from open pedestrian flow, leading to problems like mismatches, missed matches, broken links, and inability to renew links across days. Specifically, the shortcomings of existing solutions are mainly reflected in four aspects: First, physical constraints such as spatial accessibility are used as posterior scores rather than preconditions, resulting in incorrect associations that are physically inaccessible still participating in the calculation; second, the lack of a unified ground physical coordinate system makes it difficult to compare behavioral features such as trajectory length and walking speed under different cameras across cameras; third, it ignores the long-term stable spatial boundary behavioral preferences of individuals in open streets, such as the habit of walking along boundaries or the center; and fourth, it lacks a robust mechanism for template updates, link state maintenance, and cross-day link renewal, making it difficult to support long-term behavioral link reconstruction. Summary of the Invention

[0003] This invention aims to solve the technical problems existing in the above-mentioned related technologies, and proposes a method and system for reconstructing cross-camera behavior chains in open blocks. Without relying on real-name identity recognition, it can achieve anonymous long-cycle behavior chain reconstruction across cameras, time windows and days by integrating functional space hard constraints, monocular absolute scale uniformity and spatial boundary behavior preference stability weighting mechanism, which significantly reduces the risk of false association and improves the continuity and interpretability of the chain.

[0004] On the one hand, the present invention provides a method for reconstructing cross-camera behavior chains in open blocks, the method comprising the following steps: Step S100: Construct a functional space structure model in the target open block, and generate a spatial unit reachability matrix and an adaptive spatial unit transition time window based on the functional space structure model; Step S200: Acquire image data of the target behavior segment, establish a monocular absolute scale mapping relationship based on camera parameters and ground plane constraints, and restore the reference point trajectory to the absolute ground physical trajectory; Step S300: Extract absolute-scale behavioral features and behavioral preference anchor points from the absolute ground physical trajectory by combining the spatial boundary reference system; Step S400: Perform pre-rejection screening on cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window; Step S500: Combining the absolute scale behavioral features and behavioral preference anchors, construct an association cost function that includes geometric scale difference terms, absolute scale behavioral difference terms, and behavioral preference difference terms. Calculate preference stability based on the historical preferences of the anonymous behavioral chain, dynamically adjust the weights of the behavioral preference difference terms accordingly, and perform anonymous behavioral chain inheritance based on the association cost. Step S600: Perform template update, interruption recovery, and cross-day chain maintenance on the anonymous behavior chain to obtain an anonymous long-cycle behavior chain.

[0005] Furthermore, in step S100, the functional space structure model divides the open block into multiple types of functional space units, which include at least one or more of the following: main street section, side alley road section, square node, arcade corridor section, scenic spot entrance front area, commercial interface front area, transportation connection area, pedestrian gathering and dispersal area, parking connection area, and waterfront walkway section. The functional space structure model further records the one-way passage attributes, open time period attributes, temporary closure status attributes, access control restriction attributes, high congestion restriction attributes, node dwell attributes, and passable width attributes of each functional space unit.

[0006] Further, in step S100, the process of establishing the adaptive spatial unit transfer time window specifically includes: Obtain the shortest passable path length between spatial units, the historical average absolute step speed of anonymous behavior chains, the congestion coefficient of the current time period, the path type coefficient, the turning complexity coefficient, and the slope influence coefficient; The basic transfer time is calculated based on the shortest passable path length and the historical average absolute step speed of the anonymous behavior chain. The basic transfer time is then corrected using the current time period congestion coefficient, path type coefficient, turning complexity coefficient, and slope influence coefficient to determine the upper and lower bounds of the adaptive spatial unit transfer time window.

[0007] Furthermore, the reference points include one or more of the following: key foot points, center points of both feet, projection points of the human body on the ground, and center points of the bottom of the detection frame; The absolute scale behavioral characteristics are a set of features with physical dimensions calculated based on the absolute ground physical trajectory. The set of features includes one or more of the following: total trajectory length, average absolute step speed, absolute displacement amplitude, node dwell time, activity radius, lateral offset amplitude, directional change rate, and area access order. The behavioral preference anchors include edge preference anchors, center preference anchors, follower preference anchors, and avoidance preference anchors.

[0008] Furthermore, the spatial boundary reference system is used to define the relative positional relationships within the linear passage space. For any linear passage space, the spatial boundary reference system includes at least the left boundary line, the right boundary line, the center line, the passable width value, and the main passage direction vector. In step S300, the absolute-scale behavioral features and behavioral preference anchor points are extracted from the absolute ground physical trajectory by combining the spatial boundary reference system, specifically including: The absolute scale behavior characteristics with physical dimensions are calculated based on the absolute ground physical trajectory. Calculate the vertical distance between the sampling point on the target trajectory and the nearest boundary line. When the percentage of time the vertical distance is less than a preset edge threshold exceeds a first preset proportion, generate the edge preference anchor point. Calculate the degree of deviation between the sampling points on the target trajectory and the center line. When the percentage of time during which the degree of deviation is less than a preset center threshold exceeds a second preset proportion, generate the centering preference anchor point. The following preference anchor and the avoidance preference anchor are generated based on the relative speed and distance changes between the target and the trajectories of other pedestrians in the surrounding area.

[0009] Further, in step S400, a pre-emptive hard rejection screening is performed on cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window, specifically including the judgment of at least one of the following situations: Determine whether the functional space unit to which the candidate behavior segment belongs is marked as unreachable in the space unit reachability matrix; if so, reject it directly. Determine whether the time difference between the occurrence time of the candidate behavior fragment and the last occurrence time of the anonymous behavior chain is less than the lower bound of the adaptive spatial unit transfer time window; if so, reject it directly. Determine whether the time difference is greater than the upper bound of the adaptive spatial unit transfer time window; if so, reject the request directly. Determine whether the functional space unit corresponding to the current time period is in an impassable state, or whether the candidate migration direction conflicts with the one-way passage attribute; if so, reject it directly.

[0010] Further, in step S500, the preference stability is calculated based on the historical preferences of the anonymous behavioral chain, and the weight of the behavioral preference difference item is dynamically adjusted accordingly, specifically including: The frequency of occurrence of various behavioral preference anchor points in the anonymous behavioral chain is statistically analyzed within a preset historical time window, and the multi-window statistical variance and historical sample dispersion are calculated. The preference stability is determined based on the multi-window statistical variance and the historical sample dispersion, wherein the smaller the values ​​of the multi-window statistical variance and the historical sample dispersion, the higher the corresponding preference stability. If the preference stability is higher than a preset stability threshold, then the weight coefficient of the behavioral preference difference term is increased in the association cost function; If the stability of the preference is lower than the preset fluctuation threshold, the weight coefficient of the behavioral preference difference item is reduced.

[0011] Further, in step S600, performing a template update on the anonymous behavior chain specifically includes: A filtering algorithm is used to update the absolute scale behavior template of the anonymous behavior chain to adapt to the gradual change in the target step speed. The behavioral preference template is incrementally updated based on the latest behavioral preference anchor points, and the regional access template and time rhythm template are updated simultaneously. Based on the historical score fluctuations of the association cost function, the chain state confidence is updated. When the chain state confidence falls below a preset threshold, an alarm is triggered or tracking is terminated.

[0012] Furthermore, in step S600, the interruption recovery includes short-term interruption recovery, cross-time period interruption recovery, and cross-day reconnection recovery; During interrupt recovery, resuming the chain is only allowed if the candidate behavior fragment meets the following conditions simultaneously: The occurrence time of the candidate behavior fragment falls within the preset recovery time window; The spatial unit where the candidate behavior segment is located satisfies spatial migration consistency with the spatial unit before the interruption. The deviation between the absolute scale behavioral characteristics of the candidate behavioral fragment and the anonymous behavioral chain historical template is within the preset absolute scale behavioral deviation constraint range. The drift between the behavioral preference anchor point of the candidate behavioral fragment and the historical preference of the anonymous behavioral chain is within the preset behavioral preference drift constraint.

[0013] On the other hand, this application provides a cross-camera behavior chain reconstruction system for open blocks, used to perform the cross-camera behavior chain reconstruction method for open blocks as described above.

[0014] The beneficial effects of this invention are as follows: This application provides a method for reconstructing cross-camera behavior chains in open blocks. This method constructs a functional spatial structure model of the target open block and generates a spatial unit reachability matrix and an adaptive transfer time window. It then establishes a monocular absolute scale mapping based on camera parameters to recover the target's absolute ground physical trajectory, thereby extracting dimensional absolute scale behavioral features under a unified physical coordinate system. By introducing a spatial boundary reference system to identify behavioral preference anchor points, and dynamically adjusting the weight of the difference term in the association cost function based on the stability of historical preferences, coupled with a pre-emptive hard rejection screening mechanism to eliminate physically unreachable candidate segments, the method ultimately achieves template updating, interruption recovery, and cross-day chain maintenance for anonymous behavior chains. This technical solution effectively reduces the risk of false associations in cross-camera tracking of open blocks and significantly improves the stability, continuity, and interpretability of long-cycle behavior chain reconstruction. This application also provides a corresponding system, whose beneficial effects are the same as those of the above method, and will not be elaborated upon here.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0017] Figure 1 This is a flowchart of the method for reconstructing cross-camera behavior chains in open blocks provided in this application; Figure 2 This is a schematic diagram illustrating the location and tracking of targets within the coverage area of ​​multiple cameras, as provided in this application. Figure 3 This is a schematic diagram of continuous target tracking and trajectory inheritance across cameras provided in this application; Figure 4 This is a schematic diagram of a monocular camera used in this application to perform real-time tracking and geospatial mapping of a target; Figure 5 This is a structural diagram of the cross-camera behavior chain reconstruction system for open blocks provided in this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.

[0020] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0021] 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 application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0022] In open street scenarios, due to the open space, discrete camera distribution, and repeated appearance of targets across cameras, existing technologies typically rely on appearance features, re-identification features, or coarse time windows for cross-camera association. However, these methods often struggle to effectively address issues such as occlusion, lighting changes, clothing variations, and disturbances from open pedestrian flow, leading to problems like mismatches, missed matches, broken links, and inability to renew links across days. Specifically, the shortcomings of existing solutions are mainly reflected in four aspects: First, physical constraints such as spatial accessibility are used as posterior scores rather than preconditions, resulting in incorrect associations that are physically inaccessible still participating in the calculation; second, the lack of unified ground physical coordinates makes it difficult to compare behavioral features such as trajectory length and walking speed under different cameras across cameras; third, it ignores the long-term stable spatial boundary behavioral preferences of individuals in open streets, such as the habit of walking along boundaries or the center; and fourth, it lacks a robust mechanism for template updates, link state maintenance, and cross-day link renewal, making it difficult to support long-term behavioral link reconstruction.

[0023] Existing technologies mainly include cross-camera matching based on appearance features or re-identification features, candidate reduction based on camera topology and coarse time windows, short-time trajectory stitching based on pixel trajectories and local motion features, and cross-video continuous tracking based on video registration or geographic mapping. These solutions have significant shortcomings in open urban areas. First, open urban areas have clear spatial organization structures and traffic rules, such as reachability relationships, one-way traffic restrictions, open hours, and congestion restrictions. These factors directly determine whether candidate migration is physically feasible, and existing technologies usually treat them as posterior scoring factors rather than pre-existing hard thresholds.

[0024] Secondly, target trajectories under different cameras are mostly at the pixel scale, local scale, or relative scale, making it difficult to make cross-camera comparisons of trajectory length, absolute walking speed, dwell time, and activity radius. Furthermore, individuals in open streets typically exhibit relatively stable spatial boundary behavior preferences, such as walking along boundaries, walking along the central area, following targets ahead, or avoiding pedestrian obstacles in advance. However, existing technologies are insufficient in extracting and utilizing these long-term stable preferences.

[0025] Finally, most solutions are only applicable to short-term associations at the minute or hour level, lacking mechanisms for template updates, chain state maintenance, interruption recovery, and cross-day chain continuation. Therefore, a new technical solution is still needed to enable cross-mirror candidates to be screened by pre-emptive hard constraints of spatial accessibility and reasonable transfer time before entering the association cost calculation. Then, anonymous behavioral chain inheritance should be performed under a unified ground physical scale, combined with long-term stable spatial boundary behavioral preferences, and further support should be provided for interruption recovery and cross-day chain continuation maintenance.

[0026] To address the aforementioned issues, this application provides a method and system for reconstructing cross-camera behavior chains in open blocks, constructing a cross-camera behavior chain reconstruction framework that integrates physical spatial constraints, absolute scale uniformity, and behavioral preference stability weighting. First, this method constructs a functional spatial structure model in the target open block, generating a spatial unit reachability matrix and an adaptive transition time window as pre-existing hard constraints for cross-camera association, effectively eliminating erroneous candidates that are physically inaccessible. Second, by establishing a monocular absolute scale mapping relationship between camera parameters and ground plane constraints, the trajectories of reference points scattered across different viewpoints are uniformly restored to ground physical trajectories with real physical dimensions, achieving comparability of behavioral features across cameras.

[0027] Furthermore, a spatial boundary reference system is introduced to extract stable behavioral preference anchor points exhibited by individuals during long-term travel. Preference stability is calculated based on historical preferences within the anonymous behavioral chain, which is used to dynamically adjust the weight of the behavioral preference difference term in the association cost function. Finally, an association cost is constructed by combining multi-dimensional difference terms such as geometric proportion, absolute scale behavior, and behavioral preferences. Through template updates, interruption recovery, and cross-day chain continuation mechanisms, a long-term, highly continuous, and strongly interpretable cross-camera behavioral chain reconstruction of anonymous targets is achieved.

[0028] First, the cross-camera behavior chain reconstruction method for open blocks provided in this application will be described in detail below with reference to the accompanying drawings. It should be noted that this invention belongs to the upper-level behavior chain reconstruction architecture layer, mainly solving the problems of trajectory connection and identity preservation across cameras and time periods. This layer provides basic long-term behavior chain data for the downstream lower-level state interpretation and risk feedback layer.

[0029] Reference Figure 1The implementation process of the open street cross-camera behavior chain reconstruction method provided in this application embodiment includes, but is not limited to, the following steps.

[0030] Step S100: Construct a functional space structure model in the target open block, and generate a spatial unit reachability matrix and an adaptive spatial unit transition time window based on the functional space structure model.

[0031] In step S100, a priori constraint framework of the physical world is established for the subsequent reconstruction of cross-camera behavior chains. Open blocks are not unordered free spaces; their internal roads, entrances and exits, medians, and functional zoning together constitute a functional spatial structure with specific topological relationships. By constructing this model, the system can divide the physical space into spatial units with clear accessibility attributes, and based on this, deduce whether any two units are physically reachable.

[0032] Simultaneously, by combining the geographical scale of the blocks, traffic rules, and traffic flow characteristics at different times, the system can calculate the reasonable time range required for a target to move from one unit to another, i.e., the adaptive transfer time window. This preliminary modeling work transforms the originally ambiguous cross-view association problem into an ordered search problem under clear spatial and temporal constraints, fundamentally eliminating a large number of spurious associations that are physically impossible, and laying a solid foundation for the efficient and accurate operation of the entire system.

[0033] Step S200: Acquire image data of the target behavior segment, establish a monocular absolute scale mapping relationship based on camera parameters and ground plane constraints, and restore the reference point trajectory to the absolute ground physical trajectory.

[0034] Step S200 addresses the problem of scale inconsistency caused by differences in viewpoints in cross-camera tracking. In multi-camera monitoring scenarios, the pixel size, movement speed, and trajectory length of the same target vary across different cameras due to their position and distance within the field of view. This pixel-based relative scale makes it impossible to directly compare cross-camera behavioral features. By utilizing the intrinsic and extrinsic parameters of the cameras, combined with the geometric constraints of the ground as a unified plane, this step can back-project the pixel trajectory in the image coordinate system onto the real-world ground coordinate system, thereby recovering the absolute physical trajectory in meters.

[0035] This process enables the target's behavioral characteristics, such as walking speed, stride length, and dwell range, to acquire real-world physical dimensions, achieving cross-camera comparability under a unified benchmark and providing a reliable data foundation for subsequent accurate behavioral feature extraction and matching.

[0036] Step S300: Extract absolute-scale behavioral features and behavioral preference anchors from the absolute ground physical trajectory by combining the spatial boundary reference frame.

[0037] In step S300, stable and recognizable long-term behavioral patterns of individuals in open blocks are mined. After obtaining a uniform absolute-scale trajectory, the system analyzes the relative relationships between the target and spatial boundaries, center lines, and surrounding pedestrians to quantify and extract specific behavioral features such as walking along the edge, walking in the center, following the target ahead, or actively avoiding obstacles. These features not only include instantaneous motion states but are also condensed into preference anchors that represent an individual's walking habits.

[0038] Step S400: Perform pre-rejection screening on cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window.

[0039] In step S400, physical feasibility verification significantly reduces the burden of subsequent complex calculations and effectively curbs the occurrence of false associations. Before performing cross-camera matching, the system first checks the spatial relationship between candidate segments: if the camera area from which the target leaves and the camera area where it appears are determined to be physically inaccessible in the functional space structure model, the candidate is directly rejected; similarly, if the target's cross-camera transfer time is much longer or shorter than the reasonable time window calculated based on the street scale, it is also rejected.

[0040] This strategy, which uses hard physical constraints as a pre-emptive threshold, can efficiently filter out the vast majority of obviously erroneous association assumptions, preventing these incorrect candidates from entering the subsequent cost calculation and matching process. This not only significantly improves the system's operational efficiency, but more importantly, it ensures the physical rationality of the association results from the source, greatly reducing the false matching rate.

[0041] Step S500: Combining absolute scale behavioral features and behavioral preference anchors, construct an association cost function that includes geometric scale difference terms, absolute scale behavioral difference terms, and behavioral preference difference terms. Calculate preference stability based on historical preferences of the anonymous behavioral chain, dynamically adjust the weights of behavioral preference difference terms accordingly, and perform anonymous behavioral chain inheritance based on the association cost.

[0042] In step S500, the optimal matching decision is achieved through multi-dimensional feature fusion. This step formalizes the cross-camera association problem as a cost minimization problem, combining absolute-scale behavioral features and behavioral preference anchors to comprehensively consider the target's appearance geometric features, motion behavior features at a unified physical scale, and stable behavioral preference features. By constructing a multi-dimensional association cost function, the system can quantitatively evaluate the similarity between candidate segments and existing anonymous behavioral chains.

[0043] Meanwhile, by analyzing historical data of anonymous behavioral chains, the system can assess the stability of these preferences over time. This description based on long-term behavioral preferences has higher robustness and discriminative power compared to appearance features that are easily affected by occlusion, lighting, and clothing changes. It provides a strong supplementary basis for cross-camera association, enabling the system to more accurately determine whether different segments belong to the same target.

[0044] Crucially, the system dynamically adjusts the weight of behavioral preference differences in the total cost based on the calculated preference stability: for targets with high preference stability, higher weights are assigned to behavioral preferences, making them the primary basis for association; conversely, their influence is reduced. This dynamic weighting mechanism enables the system to adapt to the behavioral characteristics of different targets, making more accurate and robust chain inheritance decisions, effectively addressing the complex and ever-changing tracking scenarios in open urban areas.

[0045] Step S600: Perform template update, interruption recovery, and cross-day chain maintenance on the anonymous behavior chain to obtain an anonymous long-cycle behavior chain.

[0046] Step S600 ensures the continuity, integrity, and availability of the behavior chain over a long period. In actual monitoring, the target's behavior chain may be interrupted due to prolonged obstruction, leaving the monitoring area, or camera switching failure. This step continuously updates the feature template of the behavior chain to adapt to slow changes in the target's appearance. When a link interruption is detected, the system can initiate a recovery mechanism to search for and attempt to reconnect to the target within a reasonable time and space range.

[0047] More importantly, the system possesses the capability to continue data across days, enabling it to correlate behavioral fragments of the same anonymous target on different dates, thereby constructing long-term behavioral chains spanning multiple days. This series of maintenance mechanisms allows the system to move beyond short-term trajectory splicing and provide continuous behavioral data with long-term reference value for higher-level applications such as security early warning, passenger flow analysis, and urban planning.

[0048] It should be noted that this invention belongs to the upper-level behavior chain reconstruction architecture layer, and its output anonymous chain window object serves as the input boundary of the downstream lower-level state interpretation and risk feedback layer, forming a cascade relationship between macroscopic reconstruction and microscopic interpretation.

[0049] Furthermore, step S600 also includes receiving the structured state interpretation results output by the downstream system. When a risk score enhancement instruction is received from the downstream system, the system introduces a penalty term into the current association cost calculation to reduce the association probability of the candidate segment; when an abnormal behavior feature template update package is received from the downstream system, the system integrates it into the current behavior preference template or absolute scale behavior template to adapt to sudden changes in target behavior or correct tracking biases. Through this interface, the system achieves real-time response to downstream feedback information, completing closed-loop control from behavior chain reconstruction to risk-driven optimization.

[0050] In some embodiments of this application, in step S100, the functional space structure model divides the open block into multiple types of functional space units. The functional space units include at least one or more of the following: main street sections, side alley sections, plaza nodes, arcade corridor sections, scenic spot entrance areas, commercial interface areas, transportation connection areas, pedestrian gathering and dispersal areas, parking connection areas, and waterfront walkway sections.

[0051] Specifically, the open street blocks are divided into various types of functional space units, and the physically continuous and complex urban street scene is deconstructed into discrete semantic modules that can be understood and computed by computers. By defining specific functional attributes such as main street sections, side alley sections, square nodes, arcade connecting sections, scenic spot entrance areas, commercial interface areas, transportation connection areas, pedestrian gathering and dispersal areas, parking connection areas, and waterfront walkway sections, the system can establish a refined spatial cognition system.

[0052] This division is not merely a simple geographical partitioning, but a deep modeling of traffic patterns, pedestrian density characteristics, and target behavior patterns in different areas. For example, main street sections typically represent linear trajectories of high-speed traffic, while commercial areas or pedestrian gathering and dispersal zones are often characterized by slow-moving wandering, frequent stops, or irregular turns by the target.

[0053] These significantly different regions are defined as independent functional space units, enabling subsequent algorithms to adaptively adjust the expectations and judgment criteria for the target's motion state based on the specific environmental type in which the target is located. This greatly improves the accuracy and logic of behavior analysis in complex open street environments.

[0054] In some embodiments of this application, the functional space structure model further records the one-way passage attributes, open time period attributes, temporary closed state attributes, access control restriction attributes, high congestion restriction attributes, node dwell attributes, and passable width attributes of each functional space unit. This introduces multi-dimensional physical rules and dynamic environmental constraints for the reconstruction of cross-camera behavioral chains.

[0055] These refined attributes upgrade the originally static spatial division into a dynamic road network model with rich semantic information. For example, the one-way traffic attribute can directly determine the legitimacy of the target's movement direction, effectively filtering out false associations of reverse movement; the open time period attribute and the temporary closure status attribute introduce hard constraints in the time dimension, ensuring that the target's trajectory transfer conforms to the actual operating status of the venue at a specific time; the access control restriction attribute and the high congestion restriction attribute reflect the access control and carrying capacity threshold of personnel in a specific area, which helps to identify abnormal intrusions or avoid high-risk congestion areas.

[0056] Meanwhile, the passable width attribute, combined with the node dwell attribute, can provide the system with an expected basis for the efficiency of target passage and the reasonable dwelling behavior at specific nodes (such as the entrance area of ​​a scenic spot or a gathering and dispersing area for crowds). By integrating these attributes into the model, the system can accurately eliminate candidate segments that violate traffic rules, spatiotemporal logic, or physical limitations in the pre-screening stage, thereby greatly improving the robustness and real-world interpretability of long-cycle anonymous behavior chain reconstruction.

[0057] In some embodiments of this application, step S100, the process of establishing the adaptive spatial unit transfer time window specifically includes: Step S110: Obtain the shortest passable path length between spatial units, the historical average absolute step speed of anonymous behavior chains, the congestion coefficient of the current time period, the path type coefficient, the turning complexity coefficient, and the slope influence coefficient.

[0058] In step S110, a multi-dimensional data perception foundation is established for constructing an adaptive time-constrained model, breaking through the limitations of traditional single distance or fixed speed, and comprehensively capturing key variables affecting pedestrian movement efficiency. The system establishes a benchmark for physical distance by obtaining the shortest passable path length between spatial units; it grasps the travel habits of target individuals through the historical average absolute pace of anonymous behavioral chains; and it quantifies the dynamic interference of environmental factors on travel speed through the current time period's congestion coefficient, path type coefficient, turning complexity coefficient, and slope influence coefficient.

[0059] The collection of this series of parameters organically combines the static physical structure of the open street with the dynamic state of pedestrian flow, providing complete and necessary data support for calculating the basic transfer time and making fine-grained corrections in subsequent steps. This enables the system to generate highly personalized transfer time windows for specific targets and environments.

[0060] Step S120: Calculate the basic transfer time based on the shortest passable path length and the historical average absolute step speed of the anonymous behavior chain, and correct the basic transfer time using the current time period congestion coefficient, path type coefficient, turning complexity coefficient and slope influence coefficient to determine the upper and lower bounds of the adaptive spatial unit transfer time window.

[0061] In step S120, a highly dynamic spatiotemporal constraint model that conforms to real physical laws is constructed. It abandons the crude approach of traditional methods that rely solely on fixed speeds or straight-line distances to estimate time, instead fully considering the complexity of open street environments and the variability of individual behavior. By introducing the historical average absolute pace of anonymous behavioral chains, the system can establish personalized movement benchmarks for specific targets. Furthermore, by using multi-dimensional environmental factors such as the current time-period congestion coefficient, path type coefficient, turning complexity coefficient, and slope influence coefficient to weight and correct the basic transfer time, it accurately simulates the comprehensive impact of pedestrian density, road conditions (such as the difference between main streets and side streets), turning frequency, and terrain undulations on actual pedestrian traffic efficiency in the real world.

[0062] The final determined upper and lower bounds of the adaptive time window not only provide a highly targeted pre-rejection screening basis for cross-camera candidate behavior segments, effectively eliminating false associations that are physically impossible to achieve, but also greatly improve the accuracy and robustness of the system in reconstructing long-cycle behavior chains in complex and ever-changing environments.

[0063] In some embodiments of this application, the reference points include one or more of the following: foot key points, the center points of both feet, the projection point of the human body on the ground, and the center point of the bottom of the detection frame. In this way, the actual contact position between the target and the ground can be accurately locked, thereby ensuring the accuracy of the monocular absolute scale mapping relationship.

[0064] Specifically, in video-based measurement of human walking parameters and trajectory recovery, the visual center of the human body often shifts with movements (such as arm swinging and bending over) due to the influence of perspective projection and changes in human posture, failing to represent the true geographical location. The aforementioned reference points are all closely related to the physical contact surface between the human body and the ground. For example, foot keypoints extracted through deep learning can accurately reflect the contact state between the sole of the foot and the ground, while the center point at the bottom of the detection box provides a stable pixel-level reference.

[0065] By selecting these points as references, we can effectively eliminate height errors caused by differences in human height, limb swing, or camera tilt angle. This allows us to more realistically project the two-dimensional coordinates on the image plane into a unified ground physical coordinate system, laying a solid data foundation for the subsequent extraction of high-precision absolute ground physical trajectories.

[0066] In some embodiments of this application, the absolute scale behavior features are a set of features with physical dimensions calculated based on the absolute ground physical trajectory. The feature set includes one or more of the following: total trajectory length, average absolute step speed, absolute displacement amplitude, node dwell time, activity radius, lateral offset amplitude, directional change rate, and area access order.

[0067] Specifically, by introducing the total trajectory length and average absolute pace, the system can accurately characterize the target's movement efficiency and energy consumption level in open blocks; while the absolute displacement amplitude, activity radius, and lateral offset objectively reflect the target's activity range and its dependence on spatial boundaries during passage. Meanwhile, node dwell time records the target's lingering habits in specific functional areas (such as commercial areas or tourist attraction entrances), the rate of directional change quantifies the tortuous and hesitant characteristics of its movement route, and the area visit sequence reveals the target's flow logic between different functional spatial units.

[0068] These features, which include real physical units such as meters and seconds, completely eliminate the scale inconsistency problem caused by differences in installation height, focal length, and viewing angle among different cameras. This allows cross-camera behavior analysis to no longer be limited to local field of view, providing a stable and robust objective basis for identifying an individual's unique long-term travel habits.

[0069] In some embodiments of this application, behavioral preference anchors include edge preference anchors, center preference anchors, follower preference anchors, and avoidance preference anchors. In this way, the relatively fixed and imperceptible micro-spatial selection habits of an individual in an open street are condensed into a highly identifiable long-term behavioral fingerprint.

[0070] Specifically, unlike appearance features that are easily affected by lighting, shading, or changes in clothing, these anchor points profoundly reflect the target's subconscious logic of movement: edge preference anchor points and center preference anchor points depict the target's stable dependence on the lateral space of the road, that is, whether they tend to walk close to the building boundary or habitually occupy the central area of ​​the road; while follower preference anchor points and avoidance preference anchor points reveal the target's interaction patterns in a dynamic social environment, manifested as whether they habitually follow pedestrians in front of them or actively maintain distance to avoid obstacles in the flow of people.

[0071] By extracting these preference anchor points based on spatial boundary reference systems, the system can capture the target's stable intrinsic behavioral characteristics across different monitoring perspectives and time periods. Thus, even in the absence of clear appearance information, it can still achieve high-precision anonymous cross-view association and long-cycle behavioral chain reconstruction based on these unique spatial behavioral habits.

[0072] In some embodiments of this application, a spatial boundary reference system is used to define the relative positional relationships within a linear passage space. For any linear passage space, the spatial boundary reference system includes at least a left boundary line, a right boundary line, a center line, a passable width value, and a main passage direction vector.

[0073] Specifically, by clearly defining the left and right boundary lines (such as curbs, building setbacks, or medians) and combining them with the centerline and passable width, the system can transform trajectory data, originally based on absolute geographic coordinates, into the target's lateral offset ratio and longitudinal movement status relative to the current road boundary. This method of describing relative position not only effectively eliminates geometric interference caused by differences in direction and curvature between different road segments, but also accurately quantifies an individual's micro-spatial choice habits when traveling (e.g., whether they prefer to walk close to the edge or occupy the center of the road).

[0074] Meanwhile, the introduction of the main traffic direction vector provides a benchmark for judging the target's forward, reverse, and turning behaviors, enabling the system to extract standardized behavioral features that are not limited by macro-geographical location and have high robustness in the ever-changing street network.

[0075] In some embodiments of this application, step S300, which involves extracting absolute-scale behavioral features and behavioral preference anchor points from the absolute ground physical trajectory using a spatial boundary reference frame, specifically includes: Step S310: Calculate the absolute scale behavior characteristics with physical dimensions based on the absolute ground physical trajectory.

[0076] In step S310, the target's motion pattern is transformed from abstract pixel coordinates into quantifiable and comparable physical indicators in the real world. By extracting features with real physical units such as meters and seconds (e.g., total trajectory length, average absolute step speed, activity radius), the scale inconsistency caused by differences in installation height, focal length, and viewing angle of different cameras is completely eliminated, providing a stable and robust objective basis for identifying an individual's unique long-term travel habits.

[0077] Step S320: Calculate the vertical distance between the sampling point on the target trajectory and the nearest boundary line. When the proportion of time when the vertical distance is less than the preset edge threshold exceeds the first preset proportion, generate an edge preference anchor point.

[0078] In step S320, the individual's subconscious dependence on the horizontal space of the road is accurately captured, that is, the micro-space selection habit of whether the target tends to walk close to the building boundary, road edge or median strip is deeply depicted, thus forming a highly recognizable long-term behavioral fingerprint.

[0079] Step S330: Calculate the degree of deviation between the sampling point on the target trajectory and the center line. When the percentage of time when the degree of deviation is less than the preset center threshold exceeds the second preset ratio, generate a centering preference anchor point.

[0080] In step S330, the centering preference and edge preference complement each other, specifically used to identify pedestrians who habitually occupy the central area of ​​the road and are far from the side boundaries. This dependence on the road centerline also reflects the target's inherent spatial usage logic, which is not easily affected by the external environment, further enriching the behavioral feature dimensions of cross-view association.

[0081] Step S340: Based on the relative speed and distance changes between the target and the trajectories of other pedestrians in the surrounding area, generate follow preference anchor points and avoidance preference anchor points.

[0082] In step S340, the interaction patterns of the target in the dynamic social environment are revealed. By analyzing the speed synchronization and physical distance maintenance between the target and the nearby pedestrians, it is determined whether the target habitually follows the pedestrians in front or actively maintains distance to avoid pedestrian flow obstacles. In this way, the individual's social interaction characteristics are condensed into preference anchors that can be matched by the system.

[0083] In some embodiments of this application, step S400 involves performing a pre-emptive hard rejection screening of cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window, specifically including the determination of at least one of the following situations: In scenario (1), it is determined whether the functional space unit to which the candidate behavior segment belongs is marked as unreachable in the spatial unit reachability matrix. If so, it is directly rejected. Specifically, the physical topological constraints in the functional space structure model are used for the first round of hard filtering. By querying the pre-built spatial unit reachability matrix, the system can quickly identify remote areas that the target cannot reach from the physical road network or logical rules in the current geographical location. If the area where the candidate segment is located is marked as unreachable by the matrix, it means that the path does not exist in the real world or is permanently blocked. Therefore, it can be directly eliminated without subsequent complex spatiotemporal calculations, thereby greatly reducing the amount of invalid calculations and avoiding obvious topological errors.

[0084] In case (2), determine whether the time difference between the occurrence time of the candidate behavior fragment and the last occurrence time of the anonymous behavior chain is less than the lower bound of the adaptive spatial unit transfer time window. If so, reject directly.

[0085] Specifically, based on the speed limit principle in physics, false correlations that violate objective laws of motion are eliminated. The lower bound of the adaptive time window represents the shortest theoretical time required for a target to traverse two spatial units under ideal conditions (such as running at high speed or unobstructed passage). If the actual observed time difference is less than this lower bound, it indicates that the target has traversed a huge physical distance in an extremely short time, which is impossible in real human walking or normal movement scenarios. Therefore, the system classifies it as a false match of a different target and rejects it.

[0086] In case (3), determine whether the time difference is greater than the upper bound of the adaptive spatial unit transfer time window; if so, reject directly.

[0087] Specifically, based on the reasonable endurance and activity radius limitations of human behavior, abnormal correlations with excessively long time spans are eliminated. The upper bound of the adaptive time window comprehensively considers factors such as normal walking speed, road congestion, terrain complexity, and reasonable stops along the way, setting a maximum tolerable time for the target to complete the transfer. When the time difference exceeds this upper bound, it means that the target's movement between the two monitoring points takes too long, exceeding the logical scope of normal passage (e.g., long offline delays or detours occur midway). In this case, forcibly associating the target would disrupt the coherence and authenticity of the long-cycle behavioral chain, and therefore must be rejected.

[0088] In case (4), determine whether the functional space unit corresponding to the current time period is in an impassable state, or whether the candidate migration direction conflicts with the one-way passage attribute. If so, reject it directly.

[0089] Specifically, by introducing dynamic time rules and strict traffic control constraints, the system ensures that the reconstructed behavioral chains conform to the actual operational logic of open blocks. The system not only focuses on spatial connectivity but also verifies the validity of the time dimension (e.g., certain side streets or squares are only open during specific times) and the legality of the direction dimension (e.g., one-way streets strictly prohibit going against the flow of traffic). Through this multi-dimensional compliance check, invalid trajectories that appear reasonable in physical distance and time but actually violate access control regulations, business hours, or traffic flow rules can be effectively filtered out, thereby greatly improving the interpretability and accuracy of cross-camera tracking results in real-world scenarios.

[0090] In some embodiments of this application, step S500, calculating preference stability based on historical preferences of the anonymous behavioral chain, specifically includes: Step S510: Calculate the frequency of occurrence of various behavioral preference anchor points in the anonymous behavior chain within a preset historical time window, and calculate the multi-window statistical variance and historical sample dispersion.

[0091] In step S510, the distribution of the target's behavioral habits over a past period is quantitatively analyzed to measure the fluctuation of their behavioral patterns. By introducing the statistical concepts of variance and dispersion, the system can objectively assess whether an individual's marginal, neutral, follower, or avoidance behavioral preferences remain consistent across different time periods (i.e., multiple historical time windows). If the frequency of these preferences varies little across different time windows, it indicates that the target's common habits are highly regular and predictable; conversely, it indicates that their behavior is relatively random. This provides solid data support for subsequent assessments of the credibility of these behavioral characteristics.

[0092] Step S520: Determine preference stability based on multi-window statistical variance and historical sample dispersion, wherein the smaller the values ​​of multi-window statistical variance and historical sample dispersion, the higher the corresponding preference stability.

[0093] In step S520, the abstract behavioral fluctuation data is transformed into a dynamic weight index that can be directly used for algorithmic decision-making. Since variance and dispersion are key mathematical features characterizing the degree of data dispersion, smaller values ​​indicate a highly concentrated anchor point for the target's behavioral preferences with minimal drift, leading the system to determine extremely high preference stability. This highly stable preference is considered a strong, highly identifiable feature and will be given higher weight in subsequent cross-camera association matching. Conversely, if the values ​​are large, the system will reduce its reliance on this behavioral feature, thus achieving adaptive and precise matching based on individual behavioral characteristics.

[0094] Step S530: If the preference stability is higher than the preset stability threshold, the weight coefficient of the behavioral preference difference term is increased in the association cost function to give full play to the discriminative value of high-confidence behavioral features.

[0095] In step S530, when the system determines that the target's behavioral preferences such as edge, center, follow, or avoidance are extremely stable over a historical period (i.e., extremely low variance and dispersion), it indicates that these preferences have become highly identifiable behavioral fingerprints of the target.

[0096] Increasing the weighting coefficient of the behavioral preference difference term means that when calculating the association cost between cross-camera candidate segments and anonymous behavioral chains, the system will more rigorously examine the degree of matching between the two in terms of walking habits. This dynamic weighting mechanism can effectively amplify the habit differences between individuals, enabling pedestrians with similar appearances but vastly different walking habits to be accurately distinguished, thereby significantly improving the accuracy and robustness of cross-camera association.

[0097] In step S540, if the preference stability is lower than the preset fluctuation threshold, the weight coefficient of the behavioral preference difference item is reduced to avoid mismatches caused by over-reliance on unstable features.

[0098] In step S540, when statistical data shows that the target's behavioral preferences fluctuate drastically within the historical window (i.e., high variance and dispersion), it indicates that the target's common habits lack regularity, or that its behavioral patterns are easily affected by environmental factors, mood, and other factors, and change frequently.

[0099] In this scenario, if behavioral preferences are still used as a strong discriminative criterion, the normal behavior of the same target in different contexts may be misclassified as that of different individuals, thus severing the true behavioral chain. By reducing the weight coefficient of this discrepancy term, the system can automatically weaken its reliance on unstable features and instead rely more on relatively stable features such as geometric proportions and absolute scale behavior for association decisions, ensuring that the algorithm maintains good adaptability and reliability when facing targets with variable behaviors.

[0100] In some embodiments of this application, step S600, which involves updating the template for the anonymous behavior chain, specifically includes: Step S610: The absolute scale behavior template of the anonymous behavior chain is updated using a filtering algorithm to adapt to the gradual change in the target step speed.

[0101] In step S610, the problem of kinematic parameter drift caused by changes in physiological state, psychological intention, or environment during long-term cross-camera tracking is addressed. By introducing a filtering algorithm, the system can smoothly fuse newly observed trajectory data (such as current gait speed and displacement amplitude) and dynamically correct the original absolute-scale behavior template. This gradual update mechanism not only preserves the target's long-term stable macroscopic behavioral characteristics but also keenly captures short-term changes such as gait slowdown due to fatigue or gait speed increase due to rushing, thereby ensuring that the behavior template always maintains a high degree of consistency with the target's current true state and avoiding matching failures caused by template rigidity.

[0102] Step S620: Incrementally update the behavior preference template based on the latest behavior preference anchor point, and simultaneously update the region access template and time rhythm template.

[0103] In step S620, a multi-dimensional, holographic behavioral profile is constructed, enabling it to evolve over time. The system not only updates the target's dependence on spatial boundaries (e.g., shifting from a preference for the periphery to a preference for the center), but also enriches the regional access template (i.e., the target's frequently visited hotspots and their access frequency) and the temporal rhythm template (i.e., the target's activity patterns across different time periods) using newly acquired spatiotemporal data. This incremental, synchronous update strategy allows the anonymous behavioral chain to continuously absorb new behavioral evidence, constantly strengthening the understanding of the target's activity patterns. Thus, it maintains high-precision identification capabilities even when facing evolving target behavioral patterns, providing rich and vivid feature evidence for subsequent cross-camera association.

[0104] Step S630: Based on the historical score fluctuation of the association cost function, update the chain state confidence. When the chain state confidence is lower than a preset threshold, trigger an alarm or terminate tracking.

[0105] In step S630, by analyzing the fluctuations in historical association scores, the system can objectively assess the health of the current behavioral chain: if the score remains stable and excellent, it indicates high reliability of the association results; if the score fluctuates frequently and drastically or continues to decline, it suggests a potential risk of drastic changes in target appearance, conflicting behavioral characteristics, or mismatches. When the confidence level falls below the safety threshold, the system will automatically trigger an alarm mechanism, prompting manual review or directly terminating unreliable tracking tasks, thereby effectively preventing the accumulation and spread of erroneous associations and ensuring the reliability and rigor of the entire cross-camera tracking system.

[0106] In some embodiments of this application, in step S600, interruption recovery includes short-term interruption recovery, cross-time period interruption recovery, and cross-day reconnection recovery. When performing interruption recovery, reconnection is only allowed if the candidate behavior fragment simultaneously meets the following conditions, specifically including the following: (1) The appearance time of the candidate behavior fragment falls within the preset recovery time window.

[0107] Specifically, by utilizing the constraint of temporal continuity, reasonable timing for chain continuation that conforms to the laws of physical motion is selected. Cross-camera tracking in open streets often faces interruptions when the target briefly leaves the monitoring field of view. By setting a recovery time window calculated based on average walking speed and spatial distance, the system can effectively define the theoretical time range required for the target to move from the previous vanishing point to the current reappearance point.

[0108] Only when the appearance time of a candidate fragment falls within this window does it mean that the movement time of the target does not exceed the reasonable upper limit of human walking, nor does it violate the lower limit of the shortest passage time in physical space, thus eliminating false associations caused by abnormal time spans and ensuring the logical coherence of the behavioral chain on the timeline.

[0109] (2) The spatial unit where the candidate behavior segment is located satisfies the spatial migration consistency with the spatial unit before the interruption.

[0110] Specifically, geographical topological constraints are introduced to verify the physical accessibility and logical rationality of the target movement path. By analyzing the connection relationships between functional spatial units, the system determines whether the location of the candidate segment is adjacent to or reachable from the location before the interruption, thereby eliminating abnormal locations that are geographically blocked or require a long detour to reach.

[0111] This mechanism effectively utilizes the road network characteristics of open blocks, ensuring that the target's cross-view movement trajectory strictly follows the actual spatial layout, avoiding erroneous chain continuation caused by instantaneous movement between unconnected areas, and providing a solid spatial logical foundation for the reconstruction of behavioral chains.

[0112] (3) The deviation between the absolute scale behavioral characteristics of the candidate behavioral fragment and the anonymous behavioral chain historical template is within the preset absolute scale behavioral deviation constraint range.

[0113] Specifically, the consistency of the target's identity is ensured through stability verification of kinematic characteristics. Absolute-scale behavioral characteristics include biomechanical information with physical dimensions, such as the target's unique stride length, gait speed, height, and body shape. These characteristics exhibit high stability in the short term.

[0114] By setting a deviation constraint range, the system can quantitatively assess the similarity of candidate segments and historical templates in terms of movement patterns. If the deviation is too large, it indicates that there are significant differences in the walking posture or physiological characteristics of the two, and they are very likely to belong to different individuals. This condition effectively filters out interfering targets that are similar in appearance but have different movement habits, providing key biometric evidence for interruption recovery.

[0115] (4) The drift between the behavioral preference anchor point of the candidate behavioral fragment and the historical preference of the anonymous behavioral chain is within the preset behavioral preference drift constraint.

[0116] Specifically, long-established habitual fingerprints are used to perform high-level semantic verification of the target's identity. Behavioral preference anchors characterize the target's dependence on spatial boundaries, deviation habits from the center line, and deep-seated behavioral traits such as social interaction patterns. These traits are usually quite robust.

[0117] By limiting the drift amount, the system can determine whether the behavioral patterns of candidate segments continue historical habits within an acceptable fluctuation range. If the drift amount exceeds the constraint, it indicates that the target's travel strategy has abruptly changed, which may mean a change in identity. This condition ensures that the chain continuation is not only reasonable in time and space, but also consistent in behavioral logic, thereby greatly improving the accuracy and reliability of cross-camera tracking in complex scenarios.

[0118] In some embodiments of this application, in step S100, a functional space structure model is constructed in the target open block, and a spatial unit reachability matrix and an adaptive spatial unit transition time window are generated based on the functional space structure model. The specific implementation process of step S100 is as follows.

[0119] Step (1) involves performing a refined functional space unit division operation on the complex geographical environment of the target open block, discretizing the continuous, unstructured urban physical space into computer-understandable logical nodes. The system no longer treats the block as a single plane, but deconstructs it into several functional space units with independent attributes based on its actual use and traffic characteristics. This division preferably covers various typical urban functional areas such as main street sections, side street sections, plaza nodes, entrance areas of scenic spots, and waterfront walkways. By clearly defining these areas, the system can provide a semantically informative geographic coordinate system for subsequent target behavior analysis, ensuring that the location in each monitoring screen corresponds to a specific real-world function, thus laying a physical foundation for understanding pedestrian behavior patterns in different scenarios.

[0120] After completing the division of spatial units, this embodiment further establishes the spatial access connections between functional spatial units and the camera coverage mapping relationships, and constructs a graph model with functional spatial units as nodes. The topology of the road network in an open block was digitally reconstructed using graph theory, thus constructing a functional spatial structure model. Among these models, nodes... Representing the various functional space units defined above, while the edges... This represents the physical connection path between units.

[0121] By establishing this graph model, the system can clearly express the logical path from one area to another, while also specifying which graph node each camera's monitored image belongs to. This mapping transforms the cross-camera tracking problem into a path search problem on the graph. The system can use graph algorithms to quickly determine whether there is a physically connected path between two cameras, thus logically eliminating geographically impossible mismatches.

[0122] To give the graph model greater practical guidance significance, this embodiment also records dynamic attributes such as traffic direction, opening hours, access control status, congestion restrictions, and path type in detail on the edges of the graph model. By introducing spatiotemporal dynamic constraints, the static topology map has the ability to recognize traffic rules and operational status.

[0123] For example, the direction of travel attribute can distinguish between one-way and two-way streets, preventing logical errors in trajectory reversal; opening hours and access control status attributes clarify whether a path is available at a specific time, avoiding associating pedestrians with paths in closed or unauthorized areas at night; congestion restrictions and path type reflect the road's traffic efficiency and difficulty. By integrating this rich attribute information, the system can not only determine whether a path is connected when performing cross-camera trajectory association, but also assess the rationality and cost of path travel, thereby greatly improving the accuracy and robustness of long-term, cross-regional pedestrian re-identification.

[0124] Step (2): Based on the aforementioned functional space structure model, a spatial unit reachability matrix was established. This transforms the complex street network topology into a mathematical expression that computers can quickly query. Using this matrix, the system can definitively determine whether a physical connection exists between any two functional space units, i.e., determine the path from unit to unit. To unit Is it geographically reachable? If matrix elements... A value of 1 indicates that the two are connected; a value of 0 indicates that there is a physical barrier or the path is impassable. This binary logical expression provides the most basic hard constraint for cross-view tracking, and can quickly eliminate false trajectory assumptions that violate geographical topology (such as crossing buildings or traversing closed areas) in the initial stage of association calculation, thereby significantly reducing the computational redundancy and false detection rate of subsequent algorithms.

[0125] Step (3), in this embodiment, an adaptive transition time window is constructed. This is used to define the reasonable time range required for the target to move between two spatial units. This adaptive transfer time window A reasonable time interval was set for the pedestrian's movement across the camera. and Functional space units corresponding to the start and end points, respectively. This is the theoretical fastest arrival time. This is the slowest reasonable upper limit. This range is dynamically adjusted, taking into account factors such as the shortest path length, the average walking speed of pedestrians, the current level of congestion, road condition factors (whether the path is flat or staircase), and even the complexity of turning. Only when the time of the pedestrian's appearance in the next shot falls within this range does it conform to the laws of physics; otherwise, it will be judged as an invalid association to avoid false matching.

[0126] The adaptive transfer time window is not a static constant, but rather adjusts in real time based on various dynamic factors: it sets a base distance based on the shortest passable path length between spatial units, estimates the target's mobility by combining the historical average absolute pace of anonymous behavioral chains, introduces the current congestion coefficient to correct for slowdowns caused by crowd density, considers path type coefficients and turning complexity coefficients to reflect the impact of different road conditions (such as stairs and curves) on traffic efficiency, and finally combines the facility's openness status to ensure the path's temporal validity. Through this multi-dimensional joint calculation, the system can set a highly realistic "expected arrival time interval" for each cross-camera association. Only when the appearance time of a candidate trajectory falls within this window is it considered a reasonable association object, thus significantly improving the rigor and accuracy of cross-camera re-identification in terms of temporal logic.

[0127] In some embodiments of this application, in step S200, image data of the target behavior segment is acquired, a monocular absolute scale mapping relationship is established based on camera parameters and ground plane constraints, and the reference point trajectory is restored to the absolute ground physical trajectory. The specific implementation process of step S200 is as follows.

[0128] Step (1) establishes an independent and unified spatial boundary reference system for each accessible functional space unit, defining a standardized local coordinate system for monitoring areas of different shapes, thereby eliminating measurement inconsistencies caused by differences in camera viewing angles and installation locations. By establishing this reference system, the system can map the trajectory of the target under different cameras to the same set of geographic semantic rules for comparison, enabling the trajectory data that originally appeared messy on the image plane to obtain a unified measurement standard in physical space, providing a geometric basis for subsequent extraction of physically meaningful absolute scale behavioral features.

[0129] For linear passageways such as streets and corridors, this embodiment preferably adopts... As a unit The reference frame model accurately characterizes the geometric constraints and topological properties of the linear space through a quintuple of parameters: where and The left and right physical boundaries of the space were defined respectively, limiting the range of lateral movement for pedestrians; The central axis is indicated and serves as a baseline for measuring the degree of deviation of pedestrians. The effective passage width is defined. This represents the direction vector of travel. By introducing this parameterized model, the system can accurately calculate the lateral offset, longitudinal displacement along the path, and wandering behavior relative to the centerline of pedestrians during the travel process, thereby transforming the abstract trajectory points into physical quantities that describe specific behavioral patterns of pedestrians such as "walking on the side", "walking in the center", or "going against the flow".

[0130] For nonlinear and complex areas such as plaza nodes, intersections, and pedestrian gathering and dispersal areas, this embodiment establishes a reference system that includes a boundary envelope, a central region, and a set of main traffic directions to address the challenges of high freedom of movement and uncertain paths in open areas. By constructing the boundary envelope, the system clarifies the physical activity range of the area; by defining the central region, it identifies the core areas where crowds gather or cross; and by setting the set of main traffic directions, it captures the potential mainstream movement trends within the area. This unstructured reference system construction method enables the system to effectively describe the distance relationship of the target trajectory relative to the area boundary and its positional distribution relative to the central region. Thus, even in complex scenarios lacking fixed lane line constraints, it can still extract high-order behavioral features reflecting pedestrian traffic habits and spatial selection preferences.

[0131] Step (2) establishes a monocular absolute scale mapping relationship based on the camera's intrinsic and extrinsic parameters, true value of installation height, attitude parameters, and ground plane constraints. This breaks the limitation of missing depth information in monocular visual monitoring and constructs a mathematical bridge from the two-dimensional image pixel coordinate system to the three-dimensional real-world geographic coordinate system. By accurately calibrating the camera's physical parameters and using the ground as a reference plane, the system can calculate the corresponding position of any point in the image in the real physical space. This transforms pixels in the monitoring screen that only have relative positional relationships into geographic coordinates with real physical units of measurement (such as meters), providing a unified measurement benchmark for unifying the trajectories across cameras.

[0132] Step (3): For each captured target behavior segment, the system preferably selects the foot key point, the center point of both feet, the human body projection point, or the bottom center point of the detection box as reference points, and restores them to the absolute ground physical trajectory. Using the above mapping relationship, the target's motion trajectory in the video sequence is back-projected from the image plane to a unified ground physical coordinate system. In the formula... Represents the first The complete set of trajectories of a target in physical space Indicates the target is in the first place. The actual ground coordinates at that moment This represents the number of frames or sampling points contained in the trajectory segment. By selecting the point where the human body contacts the ground as a reference, the positional error caused by perspective projection is eliminated to the greatest extent, ensuring that the reconstructed trajectory can truly reflect the actual movement path of the pedestrian on the ground.

[0133] Through the above processing, this embodiment unifies the target trajectory from different cameras into the same ground physical coordinate system, eliminating the scale inconsistency problem caused by differences in viewing angle, installation height, and focal length in multi-camera systems. In traditional surveillance systems, the size and movement speed of the same person under different cameras are completely different in the image coordinate system, making direct comparison difficult. However, by unifying to an absolute ground physical coordinate system, the system ensures that trajectory data from any camera has the same physical semantics and measurement standards. This allows cross-camera correlation algorithms to directly calculate based on real physical distance, movement speed, and direction, greatly improving the accuracy and robustness of cross-camera tracking in complex open street environments.

[0134] In some embodiments of this application, step S300 involves extracting absolute-scale behavioral features and behavioral preference anchor points from the absolute ground physical trajectory using a spatial boundary reference system; the specific implementation process of step S300 is as follows.

[0135] Step (1): After obtaining the target's absolute ground physical trajectory, the system further extracts one or more absolute scale behavioral features, including trajectory length, average absolute step speed, absolute displacement amplitude, dwell scale, direction change rate, activity radius, lateral offset amplitude, node dwell time and area visit order. The original trajectory coordinate data is transformed into high-level semantic features with clear physical meaning, thereby constructing a "behavioral fingerprint" that can characterize an individual's unique walking habits.

[0136] For example, average absolute pace reflects a pedestrian's movement rhythm, the rate of change of direction characterizes the smoothness or hesitation of their walking, while activity radius and area visit sequence reveal their activity range and purpose within that area. These features, calculated based on real physical units (such as meters and meters per second), eliminate scale differences caused by camera perspectives, enabling the same target under different cameras to exhibit highly consistent behavioral patterns, providing a strong basis for subsequent cross-camera identity association.

[0137] Step (2): Based on this, this embodiment further extracts behavioral preference anchor points by combining the aforementioned established spatial boundary reference system. These anchor points include edge preference, center preference, following preference, and avoidance preference. From the perspective of spatial location selection, the system deeply explores the subconscious behavioral tendencies of pedestrians during their journey, adding a psychogeographical dimension to identity recognition. By quantifying the relative positional relationship between the target and spatial structures (such as boundaries and centerlines), the system can identify that some pedestrians habitually "walk on the side" to seek a sense of security, while others habitually "walk in the center" to pursue efficiency. This preference feature has extremely strong stability. Even when congestion or occlusion causes partial loss of trajectory, it can still serve as a reliable auxiliary feature for target differentiation and matching, significantly improving the system's anti-interference ability in complex open street environments.

[0138] For the extraction of edge preferences, this embodiment preferably determines them based on the distance from the target trajectory point to the nearest boundary, the proportion of time spent in the edge area, and the proportion of time spent in the near boundary zone. Through multi-dimensional statistical indicators, the degree of pedestrian dependence on the spatial edge area is accurately quantified.

[0139] The system not only calculates instantaneous distance but also tracks the duration and frequency of pedestrians staying on the edge of the path throughout the journey. This allows it to distinguish between pedestrians who occasionally approach the edge and those with a strong habit of "walking close to the edge." This refined measurement method makes edge preference a robust feature vector, effectively helping to identify targets that are visually difficult to distinguish but have significant differences in walking habits.

[0140] For extracting centering preferences, this embodiment preferably determines them based on the distance from the target trajectory point to the centerline or central area, the proportion of time spent in the central area, and the average lateral offset. This step aims to capture pedestrians' approaching behavior towards the central axis of the road or space. By calculating the average lateral offset of the trajectory point relative to the centerline, the system can determine whether the pedestrian's average position deviates from the center; combined with the proportion of time spent in the central area, it can further confirm whether they tend to be active in open or core areas.

[0141] This feature is particularly effective in distinguishing pedestrians walking on wide streets or squares, because pedestrians walking in the center tend to show stronger purpose and confidence, and their behavior patterns contrast sharply with those wandering on the periphery, thus providing complementary discriminative information for cross-view association.

[0142] Furthermore, when conditions permit, this embodiment further extracts following preferences and avoidance preferences, introducing interactive behavioral features between pedestrians to address the recognition challenges in high-density crowd scenarios. Following preferences are extracted by analyzing the distance maintained and speed synchronization between the target and pedestrians ahead, enabling the identification of groups walking together; avoidance preferences are extracted by analyzing the target's avoidance magnitude and path planning strategies when encountering obstacles or other pedestrians, reflecting an individual's reaction sensitivity and spatial avoidance habits.

[0143] These interactive behavioral features greatly enrich the dimensions of behavioral representation, enabling the system to not only focus on the independent movement of individuals, but also to use their relative behavioral patterns in the group for identity verification, thereby achieving higher-precision cross-camera tracking in complex dynamic environments.

[0144] In some embodiments of this application, step S400 involves performing a pre-rejection screening of cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window; the specific implementation process of step S400 is as follows.

[0145] Step (1): For any two candidate behavior segments across cameras and The system first identifies the functional space units to which they belong. and And calculate the time difference between the two segments. Establish a basic spatiotemporal index for candidate pairs. This represents the point in time when the previous segment ended. This represents the time point at which the next segment begins, the difference between the two. This refers to the duration during which the target disappears between the fields of view of the two cameras. By clearly defining the start and end points of these two segments in geographic space and their interval on the timeline, the system provides basic data support for subsequent verification of the rationality of the target's movement, ensuring that the correlation calculation is based on the correct spatiotemporal correspondence.

[0146] Step (2): Before performing complex feature similarity or association cost calculations, this embodiment prioritizes determining whether candidate pairs meet the pre-rejection hard condition. Utilizing the hard constraints of the physical world, it quickly eliminates a large number of obviously unreasonable false association assumptions with extremely low computational cost. In actual open street surveillance, there are massive amounts of trajectory segments. If deep feature comparisons are performed on each pair of segments, the computational load will be extremely large and inefficient. By introducing this "hard rejection" mechanism, the system essentially sets up a high-threshold filter, directly intercepting those matches that are fundamentally impossible to establish based on physical laws or traffic rules, thereby significantly narrowing the search range of subsequent algorithms and significantly improving the system's real-time processing capabilities and computational efficiency.

[0147] The specific hard rejection logic is based on reachable matrix elements. and time window constraints and Execution: If This indicates that the two spatial units are physically inaccessible; if the time difference... Less than minimum transfer time or greater than the maximum transfer time This indicates that the movement speed is illogical; or when the candidate migration direction conflicts with the one-way traffic attribute of the road, the system will directly refuse to associate it and will not enter the subsequent cost calculation process.

[0148] In this way, geographical topological connectivity, pedestrian kinematic limits, and traffic management rules are transformed into insurmountable mathematical criteria. For example, if it is impossible for a person to cross a distance of 500 meters in 10 seconds, then any calculation attempting to correlate these trajectories is meaningless. By strictly enforcing these hard physical constraints, the system can ensure that all candidate pairs entering the deep correlation stage are physically feasible, fundamentally eliminating the possibility of logical errors.

[0149] In some embodiments of this application, step S500 involves constructing an association cost function that includes geometric scale difference terms, absolute scale behavior difference terms, and behavior preference difference terms by combining absolute scale behavior features and behavior preference anchor points. The preference stability is calculated based on the historical preferences of the anonymous behavior chain, and the weights of the behavior preference difference terms are dynamically adjusted accordingly. Anonymous behavior chain inheritance is then performed based on the association cost. The specific implementation process of step S500 is as follows.

[0150] Step (1): For candidate behavioral segments that have passed the pre-selection hard rejection, the system establishes a comprehensive association cost function. A multi-dimensional scoring system is constructed to quantitatively assess the confidence that two cross-camera trajectory segments belong to the same target. By weighted summation of differences of different properties, the system can comprehensively compare candidate pairs from multiple perspectives, including geometric shape, physical scale, behavioral preferences, temporal logic, and historical features. The weight coefficients in the formula represent the weight of each feature in the overall judgment. This linear combination allows the system to flexibly adapt to different monitoring scenarios, ensuring that the final association decision is based on the globally optimal solution rather than a random match of a single feature.

[0151] In this cost function, Used to describe differences in geometric proportions This method describes absolute-scale behavioral differences, measuring trajectory similarity from a spatial kinematics perspective. Geometric scale differences primarily focus on the shape similarity of trajectories in a two-dimensional plane, such as turning angles and path curvature; it is robust to minor changes in viewpoint. Absolute-scale behavioral differences, on the other hand, utilize the recovered physical coordinates to precisely compare actual physical quantities such as pedestrian walking speed and displacement distance. The combination of these two methods allows the system to identify both the consistency of path shape and ensure that pedestrian movement speed conforms to their physiological characteristics, thus effectively distinguishing interference targets with similar paths but drastically different walking speeds.

[0152] In addition, in the function Used to describe differences in behavioral preference anchor points Used to describe differences in consistency of time sequence. Used to describe differences in historical template consistency, while The dynamic weighting coefficient is determined based on preference stability, incorporating high-level semantic features and a dynamic weighting mechanism to further improve the accuracy of the association. Behavioral preference differences measure the consistency of two trajectories in habits such as walking along the edge or in the center; temporal sequence consistency ensures that the timing of the target's appearance conforms to the physical transfer pattern; and historical template consistency compares the target's appearance or historical movement patterns. Specifically, the dynamic weighting coefficient... The introduction of this feature reflects the system's adaptability. When the system detects that a pedestrian's walking preference (such as always keeping to the right) is very stable, it will dynamically increase the weight of this feature to make it play a greater role in the discrimination. Conversely, it will reduce the weight to avoid introducing noise due to unstable features, thereby achieving personalized and accurate matching of pedestrians with different behavioral patterns.

[0153] Optionally, To measure the similarity in shape between two trajectory segments, it should be invariant to scale and translation. In practice, the two trajectory segments are first... and The coordinate sequences are normalized and mapped to a uniform unit square region to eliminate absolute positional differences caused by camera position and focal length. Then, a dynamic time warping algorithm is used to calculate the cumulative distance between the two normalized trajectory sequences. Since different pedestrians may have different walking speeds, resulting in different numbers of sampling points, dynamic time warping can find the optimal matching path between the two trajectories by non-linearly stretching or compressing the time axis. Finally, the average cumulative distance of this optimal matching path is used as the geometric ratio difference. The smaller the value, the more similar the geometric shapes of the two trajectories.

[0154] Optionally, The focus is on the consistency of the target's motion attributes in physical space. Based on the recovered absolute ground physical trajectory, the system calculates multidimensional behavioral feature vectors for both segments, including physical quantities such as average absolute step speed, trajectory length, and rate of change of direction. To eliminate the influence of different feature units (e.g., meters versus meters per second), these features are first standardized to conform to a standard normal distribution. Then, the weighted Euclidean distance between the feature vectors of the two segments is calculated. For example, differences in speed may reflect individual characteristics of pedestrians better than differences in trajectory length; therefore, the speed component is given a higher weight when calculating the distance. This distance value serves as the absolute scale behavioral difference. It can effectively distinguish targets with similar path shapes but completely different motion states (such as running and walking).

[0155] Optionally, This is used to measure the consistency of spatial usage habits between two segments. As mentioned earlier, behavioral preferences are quantified into scalar values ​​such as marginal preference and central preference. During calculation, the absolute value of the difference between the two segments on each preference indicator is directly taken. For example, if segment... The marginal preference score was 0.8 (strong marginal), while the fragment If the score is 0.1 (strongly median), then the difference between the two in marginal preferences is significant. By weighted summing the absolute values ​​of the differences across all preference dimensions, we can obtain... This calculation can strongly penalize matches that have similar paths but very different walking habits (e.g., one person walks in the middle of the road while another walks close to the wall), thus significantly improving the accuracy of associations in complex intersections or square areas.

[0156] Optionally, the dynamic weight coefficients in the association cost function This coefficient is used to adjust the proportion of behavioral preference differences in the total cost. Its calculation is based on the objective... The historical stability of behavior. System statistical objective. The variance of behavioral preference indicators over a period of time is considered. If the variance is small, it indicates that the pedestrian's walking habits are very stable (e.g., consistently walking on the right), resulting in a high degree of stability. A larger value indicates a high degree of reliance on behavioral preference characteristics during association; conversely, a smaller value suggests that the pedestrian walks randomly and has fluctuating preferences. Smaller values ​​reduce the weight of behavioral preference features in association decisions, relying mainly on appearance and geometric features for matching.

[0157] Optionally, Used to constrain the physical rationality of target transfer. Its calculation is based on segment time difference. Theoretical shortest transfer time between two spatial units The relationship is typically modeled using an exponential penalty function: when... Slightly larger At that time, the cost is minimal because it is the most logical transfer; as A gradually increasing value indicates that the target spends too much time in the middle area, and the probability that it belongs to the same person decreases exponentially, leading to a rapid increase in cost. This calculation method simulates the motion decay pattern of pedestrians, ensuring that the system prioritizes associating trajectory segments with tight temporal connections.

[0158] Optionally, This primarily describes the similarity of appearance or historical motion patterns. At the visual level, a deep convolutional neural network is used to extract image feature vectors within the target detection box, and the cosine distance between the features of the first and last frames of two segments is calculated. At the motion level, the consistency of the motion trend between the current segment and previously associated trajectories can be calculated. In practice, the image appearance distance and motion trend distance are typically linearly fused. If the appearance features of two segments are extremely similar (cosine distance close to 0), then... The cost approaches zero; conversely, if the appearance difference is huge (such as completely different colors and textures), the cost increases significantly and acts as a veto.

[0159] Optionally, the weight coefficients in the association cost function Typically obtained through offline training. A training set containing a large number of labeled positive and negative sample pairs is constructed. Machine learning algorithms such as logistic regression or support vector machines, or grid search, are used to find a set of weights that maximizes the scores of positive samples and minimizes the scores of negative samples. These weights reflect the relative importance of geometric, appearance, and behavioral features for identity determination in specific scenarios (such as dense crowds or sparse intersections).

[0160] Step (2): To improve the reliability of behavioral preference features in cross-scene association, the system further calculates the preference stability of behavioral preference anchors in the anonymous chain historical template, identifying which behavioral features are long-term inherent "strong features" of the pedestrian and which are "weak features" or "occasional features" caused by environmental interference. The calculation of preference stability is preferably determined by a combination of multi-window statistical variance, historical sample dispersion, consistency score, and long and short period drift. Specifically, the system monitors the changes and fluctuations of pedestrian preference indicators (such as the degree of walking close to the edge) through a sliding time window. If the variance is small and the consistency is high in different time periods, the preference is considered stable; otherwise, if the preference changes drastically with the scene, it is considered unstable. By introducing this stability assessment mechanism, the system can intelligently adjust the association strategy: when a certain type of preference (such as always being in the center) is stable in the long term, the weight of the corresponding preference difference item in the association cost is significantly increased, making it a strong basis for judgment; while when a certain type of preference (such as occasionally being on the left) fluctuates greatly, its weight is reduced or even blocked, thereby effectively suppressing the misleading effect of occasional behavior on trajectory matching and ensuring that the association decision mainly depends on those most representative stable features.

[0161] Step (3): After completing the above cost calculation, the system compares the minimum association cost with a preset threshold. When the minimum association cost is lower than the preset threshold, it is determined that the current candidate behavior segment belongs to the same target as the existing historical trajectory, and the candidate behavior segment is inherited into the corresponding anonymous chain to achieve trajectory continuity maintenance, that is, the newly observed segment is stitched into the existing historical trajectory chain to form a long-term cross-camera trajectory. This inheritance mechanism not only completes the identity confirmation, but also provides a more complete time clue for subsequent behavior analysis. If the association cost of all candidates is higher than the threshold, it means that the current segment belongs to a new target that has not appeared before, and the system will create a new anonymous chain for it. This dynamic chain management mechanism enables the system to continuously track and manage the movement history of hundreds or thousands of different targets in an uncontrolled environment such as open blocks.

[0162] In some embodiments of this application, step S600 involves performing template updates, interruption recovery, and cross-day chain maintenance on the anonymous behavior chain to obtain an anonymous long-cycle behavior chain; the specific implementation process of step S600 is as follows.

[0163] After inheritance or creation, the system immediately performs a comprehensive template update on the anonymous chain, including updates to the absolute scale behavior template, behavior preference template, area access template, time rhythm template, and chain state confidence, to maintain the freshness and accuracy of the historical model and enable it to adapt to the slow evolution of pedestrian behavior over time.

[0164] For example, as time goes on, a pedestrian's walking speed may slow down, or their activity area may shift. By incrementally updating the template, the system can capture these changes in a timely manner. In particular, for candidate segments with low confidence (such as segments whose features are blurred due to occlusion), the system adopts a strategy of reducing the update weight or pausing the update, setting up a firewall for historical templates to prevent noise from being introduced into long-term memory due to a single incorrect association. This avoids cumulative error pollution to subsequent association tasks and ensures the robustness of the system under long-term operation.

[0165] Furthermore, to effectively address the issue of trajectory loss in surveillance scenarios caused by obstruction, camera blind spots, or targets disappearing from view for extended periods, the system does not immediately destroy an anonymous chain if no new behavioral fragment is matched within a preset time period. Instead, it marks the chain as interrupted. This step preserves the target's "historical memory," providing a contextual basis for subsequent trajectory re-identification. In real-world open street environments, targets may enter buildings or underground passages, rendering them undetectable by cameras for a period. By introducing the interruption mechanism, the system can tolerate this temporary loss of connection, preventing a complete reset of identity recognition due to the temporary disappearance of the target, thus enabling long-term cross-camera tracking.

[0166] For anonymous chains that are in an interrupted state, when new candidate segments appear, the system will only perform the chain restoration operation if the recovery time window, spatial migration consistency, absolute scale behavior deviation constraint, and behavior preference drift constraint are satisfied simultaneously. By establishing a strict verification mechanism, it is ensured that the reconnected trajectories do indeed belong to the same target.

[0167] The recovery time window limits the longest reasonable duration of target disappearance, preventing the misassociation of unrelated trajectories separated by several days. Spatial migration consistency verifies the topological feasibility of the target's movement path from the disappearance point to the reappearance point. Absolute scale behavioral bias constraints and behavioral preference drift constraints ensure that the target's motion characteristics (such as walking speed and habits) after reappearance remain consistent with the historical template before the interruption. This combination of multi-dimensional constraints greatly reduces the risk of misassociation in complex crowds, ensuring the continuity and accuracy of cross-camera trajectories.

[0168] Furthermore, the recovery mechanism described in this embodiment has a strong ability to transcend time and space, specifically including short-term interruption recovery, cross-time period interruption recovery, and cross-day continuation recovery. If the recovery operation crosses the date boundary, the system will form a cross-day continuation, breaking through the limitations of traditional trajectory tracking which is limited to a single trip or a single day, and realizing the capture of the target's long-term life patterns.

[0169] For example, if an office worker departs from the same neighborhood every morning, the system can use cross-day chaining to connect their travel trajectories over several consecutive days, forming a long-term behavioral chain. This cross-day chaining capability is crucial for analyzing the periodic activity patterns of targets (such as commuting or regular shopping), enabling the system not only to identify who it is but also to understand what they typically do, thus providing a solid data foundation for higher-level situational awareness and behavior prediction.

[0170] In some embodiments of this application, reference is made to Figure 2 , Figure 3 and Figure 4 This demonstrates the complete process of cross-camera anonymity tracking.

[0171] Figure 2 In the center, the left panel displays the real-time status of multiple monitoring points. Two cameras (e.g., "Camera 5" and "Camera 6") are selected, and their video streams are played synchronously in the "Camera Information" panel on the right. The target being detected is marked with a red box in the image. A green dot marks the "Local Target" star in the center of the map, representing the Nth local target from the second camera, with coordinates (X1, Y1) (e.g., 100.00, 200.00) and a confidence level of 0.343. Other stars are also distributed on the map, with coordinates (X2, Y2) (e.g., 150.00, 300.00) and a confidence level of 0.634, indicating that the system is performing real-time location and tracking of targets within the coverage area of ​​multiple cameras.

[0172] Figure 3 In the cross-camera inheritance section, the system continuously tracks the same pedestrian (e.g., with ID Track001) across cameras by linking the video streams of the first camera (C1) and the second camera (C2). When the target moves from the monitoring area of ​​the first camera to the monitoring area of ​​the second camera, the system seamlessly connects the trajectory through the inheritance mechanism, ensuring that the target's identity remains consistent throughout the spatial transfer. Finally, the entire path of the target moving from the starting area (e.g., a warehouse area) to the target area (e.g., an industrial area) is fully presented on the map on the right, realizing continuous anonymous tracking of individuals across cameras and regions.

[0173] Figure 4 In the monocular tracking section, the system captures images in real time through the Mth camera (Camera M), and combines its installation parameters (such as pitch angle EA, yaw angle RA, altitude H, etc.) to continuously track targets (such as cyclists) in the image for multiple frames, generating a trajectory with a timestamp (Tstart to Tend, for example: 10:00:00 to 10:00:06), and mapping the trajectory to a digital map, marking the real-time location of the target near the target building with a star.

[0174] In this way, the target motion in the two-dimensional video footage is accurately mapped to three-dimensional geographic space, realizing the transformation from pixel coordinates to geographic coordinates. This mapping not only allows the system to locate the target's specific position near the target building in real time, but also captures the target's dynamic behavior patterns (such as cycling speed and direction changes) through trajectory generation of multiple consecutive frames, providing a high-precision spatiotemporal reference for subsequent cross-camera correlation and long-term behavior analysis.

[0175] In summary, this embodiment constructs a complete closed-loop system from local microscopic tracking to global macroscopic mapping by fusing the refined parameters of a monocular camera with a cross-camera spatiotemporal correlation mechanism. This solution not only successfully solves the accuracy problem of target localization under a single viewpoint, transforming pixel motion in the video stream into geographically significant spatiotemporal trajectories, but also overcomes the limitations of physical field of view, achieving seamless identity maintenance of the target across different monitoring areas through inheritance and chain continuation technologies.

[0176] This cross-domain and cross-time-period continuous tracking capability enables the system to extract coherent behavioral logic from fragmented video data, providing highly reliable and available data support for urban security, crowd flow situation awareness, and individual behavior pattern analysis, greatly improving the practical effectiveness of intelligent monitoring systems.

[0177] In some embodiments of this application, to verify the effectiveness of the method, anonymous behavioral segments continuously captured by 25 cameras within an open street were selected as test samples, forming 500 sets of cross-camera candidate continuation tasks. The specific experimental setup is as follows: (1) The control group was set up using only the traditional method of appearance similarity and fixed time window.

[0178] (2) The experimental group was set up using the method described in this application, which combines the spatial unit reachability matrix, adaptive spatial unit transition time window, absolute scale behavioral features and behavioral preference anchor points.

[0179] Based on the above experimental setup, the experimental results are as follows: the control group produced 108 incorrect chain continuations, with a false association rate of 21.6%; the experimental group produced 42 incorrect chain continuations, with a false association rate of 8.4%.

[0180] Therefore, this application can effectively reduce the risk of false associations in open street cross-camera scenarios and significantly improve the accuracy of long-cycle behavior chain reconstruction by introducing physical reach constraints and behavioral preference stability weighting mechanisms.

[0181] Secondly, refer to Figure 5This application provides a cross-camera behavior chain reconstruction system for open blocks, including a spatial modeling module, a trajectory recovery module, a feature extraction module, a hard rejection filtering module, a chain inheritance module, and a chain maintenance module.

[0182] In some embodiments of this application, the spatial modeling module is used to construct a functional spatial structure model and establish a spatial unit reachability matrix and an adaptive spatial unit transfer time window. This module is not only responsible for abstracting the complex urban road network into spatial units containing functional attributes and topological relationships, but also clarifies the physical connectivity between different regions by establishing a spatial unit reachability matrix, and quantifies the reasonable time range required for the target to move between different spatial nodes using an adaptive spatial unit transfer time window, thereby providing rigorous geographical constraints and spatiotemporal benchmarks for subsequent behavioral analysis.

[0183] In some embodiments of this application, the trajectory recovery module is used to recover the trajectory of reference points in a target behavior segment into an absolute ground physical trajectory using a monocular absolute scale mapping relationship. By utilizing the monocular absolute scale mapping relationship, this module can accurately backproject reference points (such as foot key points or the bottom center of the detection box) in the target behavior segment to a unified ground coordinate system, generating an absolute ground physical trajectory with real physical units such as meters and seconds. This completely eliminates geometric distortion caused by differences in camera installation height, focal length, and viewing angle, ensuring the consistency of cross-camera data in terms of physical scale.

[0184] In some embodiments of this application, the feature extraction module is used to extract absolute-scale behavioral features and behavioral preference anchors by combining a spatial boundary reference system, and to calculate preference stability. This module, by combining a spatial boundary reference system, not only calculates absolute-scale behavioral features reflecting the target's movement efficiency and activity range, but also deeply characterizes the target's dependence on road edges, centerlines, and social interaction patterns, among other behavioral preference anchors. Furthermore, it calculates preference stability using statistical methods, providing a quantitative confidence basis for the system to determine which behavioral habits can serve as long-term stable identity identifiers.

[0185] In some embodiments of this application, the hard rejection filtering module is used to perform pre-processing hard rejection filtering on cross-camera candidate behavior segments based on the spatial unit reachability matrix and the transition time window. This utilizes prior knowledge to efficiently filter invalid samples, significantly reducing computational redundancy and the risk of mismatches. Based on the spatiotemporal hard constraints set by the spatial unit reachability matrix and the transition time window, this module can quickly identify and eliminate candidate behavior segments that are physically unreachable, violate temporal laws of motion (such as faster-than-light movement or excessively long durations), or violate traffic control rules. This ensures that only trajectory segments that conform to real-world logic can enter the subsequent refined correlation calculation stage.

[0186] In some embodiments of this application, the chain inheritance module is used to dynamically adjust the association cost weights based on preference stability and perform anonymous behavior chain inheritance, achieving adaptive and accurate matching based on individual behavioral characteristics. This module abandons the traditional fixed-weight matching strategy and innovatively adjusts the weight coefficients of the behavioral preference difference term in the association cost function according to the target's preference stability: higher preference weights are assigned to targets with extremely stable behavioral habits to enhance discriminative power, while the weights are appropriately reduced for targets with variable behaviors to prevent misjudgment. This achieves highly robust automatic inheritance of anonymous behavior chains in complex and ever-changing open street environments.

[0187] In some embodiments of this application, the chain maintenance module is used to perform template updates, interruption recovery, and cross-day chain continuation maintenance on the anonymous behavior chain, ensuring that the behavior chain remains fresh and accurate in the face of environmental changes and observation interruptions. This module not only enables the behavior template to smoothly adapt to gradual changes in target pace and habit fine-tuning through filtering algorithms and incremental update mechanisms, but also establishes a robust interruption recovery and cross-day chain continuation mechanism. When the target briefly leaves the field of view or appears across different dates, it can be retrieved through multi-dimensional spatiotemporal and behavioral verification, effectively preventing the breakage and drift of the tracking chain.

[0188] In summary, the method and system for reconstructing cross-camera behavior chains in open blocks provided in this application have the following technical effects.

[0189] This technical solution successfully solves the challenge of cross-camera tracking in complex open street environments by constructing a functional space structure model and a monocular absolute scale mapping relationship, combined with multi-dimensional feature extraction and dynamic weight adjustment mechanisms. The solution not only achieves accurate conversion from pixel coordinates to "geographic coordinates" and establishes a unified spatiotemporal benchmark with physical dimensions, but also significantly improves the system's anti-interference capability and matching accuracy through pre-emptive hard rejection screening and adaptive association cost calculation based on preference stability.

[0190] This technological system demonstrates exceptional robustness and intelligence in practical applications, effectively addressing challenges such as occlusion, lighting variations, and dense crowds, enabling continuous long-term, cross-time-period behavior tracking and identity verification. Its interruption recovery and cross-day reconnection capabilities overcome the spatiotemporal limitations of traditional monitoring, providing highly reliable and available data support and core algorithm guarantees for urban security situational awareness, individual behavior pattern analysis, and smart city construction, significantly enhancing the practical effectiveness and application value of intelligent monitoring systems.

[0191] It should be noted that in all specific embodiments of this application, all data processing activities related to user identity or personal characteristics, such as user information, user behavior data, historical data, and location information, will be conducted in accordance with the principles of legality, legitimacy, and necessity. All data collection, use, storage, and processing will be subject to compliance with applicable national and regional laws, regulations, and industry standards, and informed consent from users will be obtained in a clear and explicit manner before processing. For the processing of sensitive personal information, separate consent from users will be obtained through prominent means such as pop-up prompts and independent confirmation pages. If any processing conflicts with laws and regulations, the laws and regulations will prevail, and necessary data processing will only be carried out within the scope permitted by laws and regulations, ensuring that all data-based applications, analyses, and technical implementations are conducted within the scope permitted by laws and regulations.

[0192] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0193] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.

[0194] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part 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 programs 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 of the various embodiments of this invention. 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.

[0195] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program 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 a program for use by or in conjunction with a program execution system, apparatus, or device.

[0196] More specific examples (a non-exhaustive list) of computer-readable media 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). Additionally, computer-readable media can even be paper or other suitable media on which programs can be printed, for example, by optically scanning the paper or other media, then editing, interpreting, or, if necessary, processing it in a suitable manner to obtain the program electronically, and then storing it in computer memory.

[0197] 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 program 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.

[0198] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0199] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0200] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for reconstructing cross-camera behavior chains in open street blocks, characterized in that, The method includes the following steps: Step S100: Construct a functional space structure model in the target open block, and generate a spatial unit reachability matrix and an adaptive spatial unit transition time window based on the functional space structure model; Step S200: Acquire image data of the target behavior segment, establish a monocular absolute scale mapping relationship based on camera parameters and ground plane constraints, and restore the reference point trajectory to the absolute ground physical trajectory; Step S300: Extract absolute-scale behavioral features and behavioral preference anchor points from the absolute ground physical trajectory by combining the spatial boundary reference system; Step S400: Perform pre-rejection screening on cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window; Step S500: Combining the absolute scale behavioral features and behavioral preference anchors, construct an association cost function that includes geometric scale difference terms, absolute scale behavioral difference terms, and behavioral preference difference terms. Calculate preference stability based on the historical preferences of the anonymous behavioral chain, dynamically adjust the weights of the behavioral preference difference terms accordingly, and perform anonymous behavioral chain inheritance based on the association cost. Step S600: Perform template update, interruption recovery, and cross-day chain maintenance on the anonymous behavior chain to obtain an anonymous long-cycle behavior chain.

2. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, In step S100, the functional space structure model divides the open block into multiple types of functional space units. The functional space units include at least one or more of the following: main street section, side alley road section, square node, arcade corridor section, scenic spot entrance front area, commercial interface front area, transportation connection area, pedestrian gathering and dispersal area, parking connection area, and waterfront walkway section. The functional space structure model further records the one-way passage attributes, open time period attributes, temporary closure status attributes, access control restriction attributes, high congestion restriction attributes, node dwell attributes, and passable width attributes of each functional space unit.

3. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, In step S100, the process of establishing the adaptive spatial unit transfer time window specifically includes: Obtain the shortest passable path length between spatial units, the historical average absolute step speed of anonymous behavior chains, the congestion coefficient of the current time period, the path type coefficient, the turning complexity coefficient, and the slope influence coefficient; The basic transfer time is calculated based on the shortest passable path length and the historical average absolute step speed of the anonymous behavior chain. The basic transfer time is then corrected using the current time period congestion coefficient, path type coefficient, turning complexity coefficient, and slope influence coefficient to determine the upper and lower bounds of the adaptive spatial unit transfer time window.

4. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, The reference points include one or more of the following: key foot points, center points of both feet, projection points of the human body on the ground, and the center point of the bottom of the detection frame; The absolute scale behavioral characteristics are a set of features with physical dimensions calculated based on the absolute ground physical trajectory. The set of features includes one or more of the following: total trajectory length, average absolute step speed, absolute displacement amplitude, node dwell time, activity radius, lateral offset amplitude, directional change rate, and area access order. The behavioral preference anchors include edge preference anchors, center preference anchors, follower preference anchors, and avoidance preference anchors.

5. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 4, characterized in that, The spatial boundary reference system is used to define the relative positional relationships within a linear passage space. For any linear passage space, the spatial boundary reference system includes at least a left boundary line, a right boundary line, a center line, a passable width value, and a main passage direction vector. In step S300, the absolute-scale behavioral features and behavioral preference anchor points are extracted from the absolute ground physical trajectory by combining the spatial boundary reference system, specifically including: The absolute scale behavior characteristics with physical dimensions are calculated based on the absolute ground physical trajectory. Calculate the vertical distance between the sampling point on the target trajectory and the nearest boundary line. When the percentage of time the vertical distance is less than a preset edge threshold exceeds a first preset proportion, generate the edge preference anchor point. Calculate the degree of deviation between the sampling points on the target trajectory and the center line. When the percentage of time during which the degree of deviation is less than a preset center threshold exceeds a second preset proportion, generate the centering preference anchor point. The following preference anchor and the avoidance preference anchor are generated based on the relative speed and distance changes between the target and the trajectories of other pedestrians in the surrounding area.

6. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, In step S400, a pre-emptive hard rejection screening is performed on cross-camera candidate behavior segments based on the spatial unit reachability matrix and the adaptive spatial unit transition time window, specifically including the judgment of at least one of the following situations: Determine whether the functional space unit to which the candidate behavior segment belongs is marked as unreachable in the space unit reachability matrix; if so, reject it directly. Determine whether the time difference between the occurrence time of the candidate behavior fragment and the last occurrence time of the anonymous behavior chain is less than the lower bound of the adaptive spatial unit transfer time window; if so, reject it directly. Determine whether the time difference is greater than the upper bound of the adaptive spatial unit transfer time window; if so, reject the request directly. Determine whether the functional space unit corresponding to the current time period is in an impassable state, or whether the candidate migration direction conflicts with the one-way passage attribute; if so, reject it directly.

7. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, In step S500, the preference stability is calculated based on the historical preferences of the anonymous behavioral chain, and the weights of the behavioral preference difference items are dynamically adjusted accordingly. Specifically, this includes: The frequency of occurrence of various behavioral preference anchor points in the anonymous behavioral chain is statistically analyzed within a preset historical time window, and the multi-window statistical variance and historical sample dispersion are calculated. The preference stability is determined based on the multi-window statistical variance and the historical sample dispersion, wherein the smaller the values ​​of the multi-window statistical variance and the historical sample dispersion, the higher the corresponding preference stability. If the preference stability is higher than a preset stability threshold, then the weight coefficient of the behavioral preference difference term is increased in the association cost function; If the stability of the preference is lower than the preset fluctuation threshold, the weight coefficient of the behavioral preference difference item is reduced.

8. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, In step S600, the template is updated for the anonymous behavior chain, specifically including: A filtering algorithm is used to update the absolute scale behavior template of the anonymous behavior chain to adapt to the gradual change in the target step speed. The behavioral preference template is incrementally updated based on the latest behavioral preference anchor points, and the regional access template and time rhythm template are updated simultaneously. Based on the historical score fluctuations of the association cost function, the chain state confidence is updated. When the chain state confidence falls below a preset threshold, an alarm is triggered or tracking is terminated.

9. The method for reconstructing cross-camera behavioral chains in open blocks according to claim 1, characterized in that, In step S600, the interruption recovery includes short-term interruption recovery, cross-time period interruption recovery, and cross-day reconnection recovery; During interrupt recovery, resuming the chain is only allowed if the candidate behavior fragment meets the following conditions simultaneously: The occurrence time of the candidate behavior fragment falls within the preset recovery time window; The spatial unit where the candidate behavior segment is located satisfies spatial migration consistency with the spatial unit before the interruption. The deviation between the absolute scale behavioral characteristics of the candidate behavioral fragment and the anonymous behavioral chain historical template is within the preset absolute scale behavioral deviation constraint range. The drift between the behavioral preference anchor point of the candidate behavioral fragment and the historical preference of the anonymous behavioral chain is within the preset behavioral preference drift constraint.

10. A cross-camera behavior chain reconstruction system for open street blocks, characterized in that, Method for reconstructing cross-camera behavior chains in open blocks as described in any one of claims 1 to 9.

Images