Baseline detection and risk feedback method for anonymous behavior chain and related device thereof

Abstract

The application provides a baseline detection and risk feedback method for an anonymous behavior chain and related equipment, and relates to the technical field of cross-camera target tracking. The technical scheme directly acquires an anonymous chain window object output by an upstream system and constructs a joint state representation vector based on an absolute ground physical trajectory sequence. According to multiple conditions, a historical stable window is screened to establish an individual historical baseline, and the interference of low-quality samples is effectively eliminated. At the same time, a freezing and recovery maintenance mechanism based on continuous deviation and back-to-stable conditions is introduced, which not only prevents the pollution of the individual historical baseline by the abnormal stage, but also flexibly adapts to the dynamic changes of the subject behavior. Finally, only when the trend state is continuous deviation and the explanation confidence meets the standard, the structured feedback is triggered, which significantly improves the reliability of risk determination, provides accurate and executable chain-level risk processing basis for the upstream system, and enhances the stability and intelligent level of the overall behavior analysis system.

Description

Technical Field

[0001] This invention relates to the field of cross-camera target tracking technology, and in particular to a baseline detection and risk feedback method and related equipment for anonymous behavioral chains. Background Technology

[0002] In open streetscapes or crowded environments with multiple cameras, while upstream anonymized long-cycle behavioral chain reconstruction systems can connect discrete segments to form anonymized chains, subsequent state interpretation and baseline maintenance remain challenging. Existing anomaly detection technologies are mostly based on general time-series data or ordinary trajectories, lacking clear input boundaries for anonymized chain window objects, and often indiscriminately include all historical windows in baseline updates, easily leading to coarse baseline sample selection and poor representativeness. Furthermore, existing methods lack continuous window linkage control mechanisms, making it difficult to stably distinguish between sustained shifts and short-term fluctuations, easily causing contamination of individual historical baselines during anomaly phases. Simultaneously, the risk feedback triggering conditions are unclear, relying heavily on single-window anomaly scores, lacking joint constraints on trend states and interpretative confidence, affecting the accuracy and feasibility of chain-level risk processing. Summary of the Invention

[0003] This invention aims to solve the technical problems existing in the above-mentioned related technologies, and proposes a baseline detection and risk feedback method and related equipment for anonymous behavior chains. It can establish a highly representative individual baseline by constructing a joint state representation vector and a historical stable window screening mechanism, and use a freeze-recovery strategy with continuous window linkage to prevent abnormal contamination. Finally, it triggers structured risk feedback based on trend state and confidence, thereby improving the stability of long-term behavior analysis of anonymous subjects and the executability of chain-level risk processing.

[0004] On the one hand, this invention provides a baseline detection and risk feedback method for anonymous behavioral chains, the method comprising the following steps: Obtain the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system. The anonymous chain window object includes at least the anonymous chain identifier, window identifier, window time range, and absolute ground physical trajectory sequence. Multiple component state quantities are constructed based on the absolute ground physical trajectory sequence in the anonymous chain window object, and the multiple component state quantities are concatenated to form a joint state representation vector. For historical window objects corresponding to the same anonymous chain identifier, historical stable windows are selected based on trajectory recovery quality, chain confidence, standardized offset of joint state representation vector, and number of continuous stable windows, and individual historical baselines are established based on the historical stable windows. The component state quantity and joint state representation vector corresponding to the current window are compared with the individual historical baseline to form a state comparison result; When the current window or a consecutive preset number of windows meet the continuous offset condition, the individual historical baseline is frozen and maintained; when the subsequent consecutive preset number of windows meet the stabilization condition, the individual historical baseline is restored and maintained. A structured state interpretation result is generated based on the state comparison result. When the trend state is continuously shifting and the confidence level of the state interpretation is higher than the feedback threshold, the state interpretation result is fed back to the upstream anonymous long-cycle behavior chain reconstruction system for chain-level risk handling.

[0005] Furthermore, the acquisition of the anonymous chain window object output by the upstream anonymous long-cycle behavior chain reconstruction system specifically includes: In open street blocks or pedestrian environments with multiple cameras discretely covered, receive the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system, which has been connected by pedestrian re-identification or cross-camera tracking technology. The anonymous chain window object also includes trajectory recovery quality information and chain confidence information.

[0006] Furthermore, the component state quantities specifically include: The average displacement state quantity is used to characterize the average movement distance within the current window, and is determined based on the average displacement magnitude of adjacent trajectory points within the current window. The average velocity state quantity is used to characterize the average movement speed within the current window, and is determined based on the average displacement per unit time within the current window. The direction change discrete state quantity is used to characterize the stability of the motion direction within the current window. It is determined based on the variance, standard deviation, or quantile dispersion of the continuous trajectory direction angle change within the current window. The stop-go switching count is used to characterize the frequency of motion pauses within the current window, and is determined based on the number of times the speed crosses the stop-go threshold within the current window. Periodic fluctuation state quantities are used to characterize the motion rhythm within the current window. They are determined based on the periodic amplitude, dominant frequency energy, or periodic fluctuation range of the velocity sequence, displacement sequence, or gait rhythm sequence within the current window.

[0007] Furthermore, the step of selecting historical stable windows for the same anonymous chain identifier based on trajectory recovery quality, chain confidence, standardized offset of the joint state representation vector, and the number of continuous stable windows specifically includes: Historical stable windows are defined as those whose trajectory recovery quality is higher than a first threshold, whose chain confidence is higher than a second threshold, whose joint state representation vector standardized offset is lower than a first offset threshold, and whose historical windows satisfy the stability condition for a consecutive preset number of times. The mean vector of the joint state representation vector of the selected historical stable window samples is calculated and used as the baseline of the joint state representation vector. Calculate the historical fluctuation range vector for the historical stable window sample, which serves as the fluctuation boundary of the individual historical baseline; The historical fluctuation range vector is composed of a standard deviation vector, a quantile interval vector, or a robust fluctuation range vector.

[0008] Furthermore, the step of comparing the component state quantities and joint state representation vector corresponding to the current window with the individual historical baseline to form a state comparison result specifically includes: Calculate the standardized offset of the current window joint state representation vector relative to the individual historical baseline, and the component standardized offset of each component state quantity relative to the corresponding component baseline. The state comparison result includes at least the standardized offset, component standardized offset, offset direction information, and continuous trend information.

[0009] Furthermore, the step of freezing and maintaining the individual historical baseline when the current window or a consecutive preset number of windows meet the continuous offset condition specifically includes: When the standardized offset of the joint state representation vector of the current window or a consecutive preset number of windows is higher than the freezing threshold, and the standardized offset of the component state quantity exceeding the preset number threshold exceeds the corresponding component threshold, baseline freezing is triggered; in the frozen state, the current window does not participate in the update of the individual historical baseline. When a subsequent preset number of windows meet the stabilization condition, the individual historical baseline is restored and maintained, specifically including: When the standardized offset of the joint state representation vector of a subsequent preset number of windows is lower than the recovery threshold, and the component state quantities exceeding the preset number threshold re-enter their respective stable intervals, baseline recovery is triggered; the recovery threshold is less than the freeze threshold, forming a hysteresis interval.

[0010] Furthermore, the step of generating a structured state interpretation result based on the state comparison result specifically includes: The main state label is generated based on the offset direction information and the dominant offset component information, and the trend state is determined based on the continuous trend information. The trend state includes continuous offset, continuous recovery, short-term fluctuation and stable maintenance. The step of feeding back the state interpretation results to the upstream anonymous long-cycle behavior chain reconstruction system specifically includes: Only when the trend state is a continuous shift and the state interpretation confidence is higher than the feedback threshold, the state interpretation result, which includes information on the dominant shift component and the risk processing participation identifier, will be fed back to the upstream system, triggering chain-level risk scoring enhancement or manual review priority ranking.

[0011] On the other hand, this application provides a baseline detection and risk feedback system for anonymous behavioral chains, characterized by steps for performing the aforementioned baseline detection and risk feedback method for anonymous behavioral chains.

[0012] On the other hand, this application provides an electronic device, including: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the processors perform the baseline detection and risk feedback method for anonymous behavior chains as described above.

[0013] On the other hand, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the baseline detection and risk feedback method for anonymous behavioral chains as described above.

[0014] The beneficial effects of this invention are as follows: This application provides a baseline detection and risk feedback method for anonymous behavioral chains. This technical solution directly obtains the anonymous chain window object output by the upstream system and constructs a joint state representation vector based on the absolute ground physical trajectory sequence, achieving seamless integration with the upstream reconstruction system and standardization of feature expression. This method selects historically stable windows based on multiple conditions such as trajectory recovery quality and chain confidence to establish individual historical baselines, effectively eliminating interference from low-quality samples and ensuring the representativeness and accuracy of the baseline. Simultaneously, the introduction of a freezing and recovery maintenance mechanism based on continuous offset and stabilization conditions not only prevents the contamination of individual historical baselines during abnormal phases but also flexibly adapts to the dynamic changes in subject behavior. Finally, by triggering structured feedback only when the trend state is a continuous offset and the explanatory confidence meets the standard, the reliability of risk assessment is significantly improved, providing the upstream system with accurate and executable chain-level risk handling basis, thereby enhancing the stability and intelligence level of the overall behavioral analysis system. This application also provides related equipment for the above method; the beneficial effects of the related equipment are the same as those of the above method and will not be elaborated 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 baseline detection and risk feedback method for anonymous behavioral chains provided in this application; Figure 2 This is a schematic diagram illustrating the principle of obtaining the anonymous chain window object provided in this application; Figure 3 This is a schematic diagram illustrating the principle of constructing component state variables provided in this application; Figure 4 This is a schematic diagram illustrating the principle of establishing an individual's historical baseline, as provided in this application. Figure 5 This is a schematic diagram illustrating the principle of baseline freezing and restoration provided in this application; Figure 6 This is a structural diagram of the baseline detection and risk feedback system for anonymous behavioral chains 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 the field of video surveillance and behavior analysis technology for open public spaces, existing technical architectures are typically divided into two main stages. The upstream stage relies on an anonymous long-cycle behavior chain reconstruction system. The core task of this system is to effectively connect discrete observation segments of the same anonymous subject in complex environments with multiple cameras, thereby forming a continuous anonymous behavior chain. This stage provides the basic data structure for subsequent analysis. However, when the data flows to the downstream stage—that is, when continuously interpreting the state of the formed behavior chain, maintaining individual historical baselines, and assessing anomaly risks—existing technologies reveal significant limitations and gaps.

[0023] Most existing anomaly detection and risk assessment technologies follow traditional processing paradigms, which typically extract features directly from general time-series data or ordinary moving object trajectories. Their core logic is to compare the extracted features with a pre-set population average template or a static historical baseline to output a risk level. This general approach exhibits several significant shortcomings when dealing with long-cycle behavioral chains with specific structures and anonymity attributes output from upstream sources.

[0024] First, existing technologies lack precise adaptation to the upstream output data structure, resulting in ambiguous input boundaries. They typically do not establish explicit input interfaces constrained by anonymous chain window objects, making data interaction with upstream systems unstable and difficult to guarantee the continuity and consistency of analysis.

[0025] Secondly, the selection process for baseline samples is too crude and lacks an effective screening mechanism. Existing methods often indiscriminately consider all historical windows when updating an individual's historical baseline. This approach easily introduces unstable or inherently anomalous samples, contaminating the individual's historical baseline, causing it to lose its due representativeness, and failing to accurately reflect the subject's normal behavioral patterns.

[0026] Furthermore, existing methods lack a dynamic control mechanism based on continuous window linkage. Due to this lack, the system struggles to effectively distinguish between persistent abnormal shifts and transient random fluctuations in behavioral characteristics. This not only may lead to false alarms about normal fluctuations, but more importantly, it cannot promptly prevent the contamination of an individual's historical baseline when an anomaly is detected, causing the baseline to drift over time and further reducing detection accuracy.

[0027] Finally, the triggering conditions for risk feedback are vague and singular. Most existing technologies determine whether to trigger feedback based solely on whether the anomaly score in a single window exceeds a threshold. This static judgment method lacks strict joint constraints on behavioral trend states and interpretive confidence levels, resulting in feedback results that often lack sufficient executability and accuracy, making it difficult to meet the high requirements of chain-level risk handling.

[0028] To address the aforementioned issues, this application provides a baseline detection and risk feedback method and related equipment for anonymous behavioral chains. This technical solution constructs a stable window screening mechanism that incorporates multi-dimensional conditions such as trajectory recovery quality and chain confidence by precisely adapting to the anonymous chain window object output by the upstream system, thereby establishing a highly representative individual historical baseline. Based on this, a baseline freezing and recovery maintenance strategy based on continuous window linkage is introduced, effectively preventing baseline contamination during abnormal phases. Finally, by combining the dual constraints of trend status and interpreted confidence to trigger structured feedback, the entire process from feature selection and baseline maintenance to risk assessment is optimized, significantly improving the stability of anonymous behavioral chain analysis and the accuracy of risk handling.

[0029] First, the baseline detection and risk feedback method for anonymous behavioral chains 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 low-level state interpretation and risk feedback layer, mainly addressing the problems of refined state interpretation, baseline maintenance, and risk assessment for long-cycle behavioral chains output from the upper-level behavioral chain reconstruction architecture layer. This invention supplements and optimizes the upper-level reconstruction architecture, forming a closed loop from reconstruction to interpretation and then to feedback.

[0030] Reference Figure 1 The implementation process of the baseline detection and risk feedback method for anonymous behavior chains provided in this application embodiment includes, but is not limited to, the following steps.

[0031] Step S100: Obtain the anonymous chain window object output by the upstream anonymous long-cycle behavior chain reconstruction system.

[0032] The anonymous chain window object includes at least an anonymous chain identifier, a window identifier, a window time range, and an absolute ground physical trajectory sequence.

[0033] In step S100, a standardized data interface was established with the upstream anonymous long-period behavior chain reconstruction system. By directly acquiring anonymous chain window objects containing anonymous chain identifiers, window identifiers, window time ranges, and absolute ground physical trajectory sequences, it was ensured that the data units processed in this application were valid fragments after being cleaned and correlated by the upstream system. This not only solved the interface instability problem caused by fuzzy input boundaries in the prior art, but also provided unified and accurate time and space constraints for subsequent feature extraction and baseline comparison, which is the foundation for realizing subsequent refined analysis.

[0034] Furthermore, to clarify the data processing basis of this application, the physical meaning of the above fields is defined as follows: the anonymous chain identifier, as the identity fingerprint output by the upstream reconstruction system, represents the logical continuity of the same moving entity in the physical world when switching between shots; the window identifier and the window time range together define the spatiotemporal slice of data processing, discretizing the continuous video stream into static units that can be computed; the absolute ground physical trajectory sequence is a set of points in the Cartesian coordinate system in meters (m) recovered by the upstream monocular mapping, each point corresponding to the instantaneous geographical location of the target on the real world ground, making the state quantities calculated subsequently have real physical interpretability; the chain confidence quantifies the possibility that the trajectory segment is affected by occlusion, lighting changes, or ID switch interference, reflecting the purity of the trajectory; the area access sequence records the target's travel intention in the macroscopic topology.

[0035] This step ensures that the data units processed in this application are valid fragments that have been cleaned and correlated by the upstream system, providing a clean data source for subsequent baseline establishment.

[0036] Step S200: Construct multiple component state variables based on the absolute ground physical trajectory sequence in the anonymous chain window object, and concatenate the multiple component state variables to form a joint state representation vector.

[0037] In step S200, the original trajectory data is transformed into a quantifiable multidimensional feature space representation. Multiple component state variables are constructed based on the absolute ground physical trajectory sequence, capable of characterizing the subject's behavioral features from different dimensions, such as movement speed, direction changes, and dwell time. These component state variables are concatenated to form a joint state representation vector, achieving a comprehensive abstraction of the subject's behavioral patterns within the window period. This high-dimensional vector representation not only preserves the spatiotemporal characteristics of the trajectory but also enhances the distinguishability of features, providing an effective data foundation for subsequent accurate comparison with individual historical baselines.

[0038] Step S300: For the historical window object corresponding to the same anonymous chain identifier, select historical stable windows based on trajectory recovery quality, chain confidence, standardized offset of joint state representation vector, and number of continuous stable windows, and establish individual historical baselines based on historical stable windows.

[0039] In step S300, for historical window objects with the same anonymous chain identifier, multiple screening conditions are introduced, including trajectory recovery quality, chain confidence, joint state representation vector normalization offset, and the number of continuous stable windows. This effectively eliminates low-quality or abnormal windows caused by occlusion, misidentification, or transient abnormal behavior. This rigorous screening mechanism ensures that the samples used to establish individual historical baselines are all stable and reliable segments of normal behavior, thereby avoiding the baseline drift problem caused by sample contamination in existing technologies, and enabling the baseline to truly reflect the subject's regular behavioral patterns.

[0040] Step S400: Compare the component state variables and joint state representation vector corresponding to the current window with the individual historical baseline to form a state comparison result.

[0041] In step S400, a quantitative comparison is made between the current behavior and historical norms. By comparing the component state quantities and joint state representation vector of the current window with the individual historical baseline established in step S300, specific state comparison results can be generated. This comparison is not merely a simple calculation of numerical differences, but a distance metric based on a multi-dimensional feature space, which can keenly capture the degree of deviation of the current behavior from historical norms. This result provides direct numerical evidence for subsequent judgments on whether the behavior is abnormal and the severity of the abnormality.

[0042] Step S500: When the current window or a consecutive preset number of windows meet the continuous offset condition, perform freeze maintenance on the individual historical baseline; when the subsequent consecutive preset number of windows meet the stabilization condition, perform recovery maintenance on the individual historical baseline.

[0043] In step S500, a dynamic baseline maintenance mechanism is introduced to address the problem of baseline contamination caused by anomalous behavior. When the current window or multiple consecutive windows meet the persistent offset condition, the system performs freeze maintenance on the individual historical baseline to prevent anomalous data from updating the baseline, thereby maintaining baseline stability. When subsequent windows continuously meet the stabilization condition, recovery maintenance is performed, allowing the baseline to relearn a new stable state. This freeze and recovery strategy based on continuous window linkage effectively distinguishes between persistent anomalous offsets and short-term random fluctuations, ensuring that the baseline can adapt to long-term changes in subject behavior while avoiding the damage to the baseline caused by short-term anomalies.

[0044] Step S600: Generate structured state interpretation results based on state comparison results, and when the trend state is continuously shifting and the confidence level of the state interpretation is higher than the feedback threshold, feed the state interpretation results back to the upstream anonymous long-cycle behavior chain reconstruction system for chain-level risk handling.

[0045] In step S600, a structured state interpretation result is generated based on the state comparison result. A dual trigger condition is set: a continuous deviation in trend state and a state interpretation confidence level higher than the feedback threshold. This ensures the accuracy and executability of the feedback result. The state interpretation result is only fed back to the upstream system when the behavioral deviation is not only continuous but also when the system has a sufficiently high confidence level in its interpretation. This rigorous feedback mechanism avoids false alarms caused by accidental fluctuations or low-confidence judgments, providing accurate and reliable intelligence support for chain-level risk handling in the upstream system, and significantly improving the intelligence level and practical value of the entire behavioral analysis system.

[0046] Furthermore, the structured state interpretation results specifically include risk scoring enhancement instructions, manual review priority indicators, and abnormal behavior feature template update packages. This feedback result is transmitted to the upper-level behavior chain reconstruction architecture layer (i.e., the upstream system) to dynamically adjust the relevant weight parameters of the upstream system, thereby achieving adaptive tracking optimization based on risk state. This forms a closed-loop cyclical relationship from upstream reconstruction, downstream detection, feedback correction, to upstream optimization.

[0047] In some embodiments of this application, step S100, obtaining the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system, specifically includes: in an open street or a pedestrian flow environment with discrete coverage by multiple cameras, receiving the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system, which has been concatenated through pedestrian re-identification or cross-camera tracking technology. The anonymous chain window object further includes trajectory recovery quality information and chain confidence information.

[0048] Specifically, in open street blocks or pedestrian environments with discrete coverage by multiple cameras, the anonymous chain window objects output by the upstream anonymous long-cycle behavior chain reconstruction system, which have been linked together by pedestrian re-identification or cross-camera tracking technology, are received. This process establishes the basic data source for analysis and its specific application scenarios.

[0049] This approach not only limits the technical solution to operation in complex environments with discontinuous camera coverage, but also ensures that the analyzed objects are complete behavioral chain segments spanning time and space by accepting data that has already been processed by pedestrian re-identification or cross-camera tracking technologies. The introduction of anonymous chain window objects containing trajectory recovery quality information and chain confidence information provides key metadata regarding data integrity and correlation reliability for subsequent processing. This allows the system to assess the credibility of the input chain at the initial stage of analysis, thereby laying a solid data foundation for building accurate individual behavioral baselines.

[0050] In some embodiments of this application, step S200 specifically includes the component state quantity: (1) The average displacement state quantity is used to characterize the average movement distance within the current window, and is determined based on the average displacement modulus of adjacent trajectory points within the current window.

[0051] Specifically, the average displacement state quantity is used to quantify the average spatial movement span of an object within the observation period. This indicator effectively reflects the object's active range and exploration ability within a specific area, and is a fundamental characteristic for distinguishing macroscopic behavioral patterns such as normal wandering, rapid traversal, or cramped loitering. By capturing changes in the average movement distance, the system can preliminarily determine whether the object's travel purpose has changed, such as from daily commuting to abnormal long-distance wandering.

[0052] (2) The average velocity state quantity is used to characterize the average moving speed within the current window, and is determined based on the average displacement per unit time within the current window.

[0053] Specifically, the mean speed state quantity is used to characterize the overall speed trend of a subject within the current window. This component can intuitively reveal whether the subject is leisurely strolling, walking normally, or running quickly, and is a key indicator for measuring the urgency and abnormality of behavior. In specific scenarios, abnormal increases or decreases in speed often directly correspond to potential risky behaviors, such as evading tracking or suspicious stakeouts, providing important speed dimension information for subsequent risk assessment.

[0054] (3) The direction change discrete state quantity is used to characterize the stability of the motion direction within the current window, and is determined based on the variance, standard deviation or quantile dispersion of the direction angle change of the continuous trajectory within the current window.

[0055] Specifically, the discrete state quantity of directional change is used to evaluate the smoothness and directional stability of the subject's movement path from multiple dimensions. This indicator can keenly detect whether the subject exhibits hesitant behavioral characteristics such as frequent swaying left and right, aimless turning, or sudden U-turns. High dispersion usually indicates that the subject is in a state of confusion, reconnaissance, or deliberately evading surveillance. By quantifying the degree of directional confusion, the system can effectively identify abnormal wandering patterns that differ from normal straight-line walking.

[0056] (4) The stop-go switching count is used to characterize the frequency of motion pauses within the current window, and is determined based on the number of times the speed crosses the stop-go threshold within the current window.

[0057] Specifically, the stop-and-go switching count is used to characterize the frequency and intermittency of pauses in the subject's movement rhythm. This component can accurately count the number of times the subject starts and stops during movement, reflecting the continuity of behavior. Frequent stop-and-go switching often suggests that the subject is observing, waiting for a specific opportunity, or being disturbed by external factors. This indicator has extremely high discriminative value for identifying suspicious behaviors with intermittent characteristics, such as following, tailing, or stationary observation.

[0058] (5) Periodic fluctuation state quantities are used to characterize the motion rhythm within the current window and are determined based on the periodic amplitude, dominant frequency energy, or periodic fluctuation range of the velocity sequence, displacement sequence, or gait rhythm sequence within the current window.

[0059] Specifically, the periodic fluctuation state quantity is used to uncover the hidden rhythmic patterns and regularities behind the subject's movement. This component aims to capture the inherent biomechanical characteristics and habitual rhythms in human movement, distinguishing regular periodic movements from chaotic random movements. By analyzing abnormal changes in movement rhythm, the system can further verify the naturalness of behavioral patterns and help determine whether there are any abnormalities such as faking or deliberately altering gait.

[0060] In some embodiments of this application, step S300 involves filtering historical stable windows for the same anonymous chain identifier based on trajectory recovery quality, chain confidence, standardized offset of joint state representation vector, and the number of continuous stable windows. This process includes the following steps.

[0061] Step S310: Select historical windows that meet the stability conditions, such as trajectory recovery quality higher than the first threshold, chain confidence higher than the second threshold, joint state representation vector standardized offset lower than the first offset threshold, and a preset number of consecutive historical windows, as historical stable windows.

[0062] In step S310, a rigorous multi-dimensional admission mechanism is established to accurately identify high-quality and normally behaving samples from massive historical data. By simultaneously setting hard indicators such as trajectory recovery quality exceeding a first threshold, chain confidence exceeding a second threshold, and joint state representation vector standardized offset below a first offset threshold, this step can effectively eliminate low-quality and noisy data caused by sensor errors, algorithm mismatches, or transient abnormal behavior of the subject. Furthermore, a predetermined number of consecutive historical windows must meet stability conditions, ensuring that the selected historical stable windows not only have reliable single-point data but also exhibit consistent stationarity over time, thus providing a clean data foundation for constructing a highly reliable individual baseline.

[0063] Step S320: Calculate the mean vector of the joint state representation vector of the selected historical stable window samples, and use it as the baseline of the joint state representation vector.

[0064] In step S320, the mathematical center of the individual's normal behavioral pattern is established, that is, a standardized reference benchmark is constructed. By calculating the mean of the joint state representation vectors of all historical stable window samples selected in step S310, an average feature vector that can represent the long-term behavioral habits of the anonymous subject is obtained. This mean vector serves as the baseline of the joint state representation vector, quantifying the central tendency of the subject's core characteristics such as displacement, velocity, direction, and rhythm under normal conditions, providing the most intuitive comparative reference for subsequent judgment on whether the current behavior has deviated.

[0065] Step S330: Calculate the historical fluctuation range vector for the historical stable window samples, which serves as the fluctuation boundary of the individual historical baseline. The historical fluctuation range vector is composed of a standard deviation vector, a quantile interval vector, or a robust fluctuation range vector.

[0066] In step S330, a reasonable fluctuation range for individual behavior is defined, giving the baseline dynamic tolerance. Relying solely on a single mean vector is insufficient to encompass the natural fluctuations in an individual's behavior under different circumstances. Therefore, this step constructs the fluctuation boundary of the individual's historical baseline by calculating the standard deviation vector, quantile interval vector, or robust fluctuation range vector of historical stable window samples. This historical fluctuation range vector clarifies the maximum permissible deviation range of various characteristics under normal conditions, enabling the system to scientifically distinguish between behavioral fine-tuning that belongs to normal physiological or psychological fluctuations and truly potentially threatening abnormal mutations during risk detection, significantly reducing the false alarm rate.

[0067] In some embodiments of this application, step S400 compares the component state variables and joint state representation vector corresponding to the current window with the individual historical baseline to form a state comparison result. Specifically, this includes calculating the standardized offset of the joint state representation vector of the current window relative to the individual historical baseline, and the component standardized offset of each component state variable relative to its corresponding component baseline. The state comparison result includes at least the standardized offset, component standardized offset, offset direction information, and continuous trend information.

[0068] Specifically, the process first calculates the standardized offset of the current window joint state representation vector relative to the individual's historical baseline, thereby comprehensively measuring the degree to which the subject's current integrated behavioral pattern deviates from its long-term habits. Simultaneously, it further refines the calculation of the standardized offsets of each component state quantity (such as velocity and direction) relative to its corresponding component baseline to identify which specific motion feature has undergone an abnormal change. This process transforms abstract behavioral trajectories into measurable numerical deviations, providing objective and accurate data support for subsequent risk assessment.

[0069] Building upon this foundation, the state comparison results generated in this step also encompass offset direction information and continuous trend information, significantly enriching the dimensions of anomaly detection. Offset direction information clearly reveals whether the current behavior is positively enhancing or negatively weakening (e.g., whether the speed is abnormally accelerating or slowing down), while continuous trend information records the evolution of this offset over time. This rich information not only helps the system determine whether an anomaly exists but also further explains how the anomaly occurs and its development trend, laying a solid foundation for ultimately generating executable, structured risk feedback.

[0070] In some embodiments of this application, in step S500, when the current window or a consecutive preset number of windows meet the continuous offset condition, the individual historical baseline is frozen and maintained, specifically including the following steps.

[0071] Step S510: When the standardized offset of the joint state representation vector of the current window or a consecutive preset number of windows is higher than the freeze threshold, and the standardized offset of the component state variables exceeding the preset number threshold exceeds the corresponding component threshold, baseline freezing is triggered. In the frozen state, the current window does not participate in updating the individual historical baseline.

[0072] In step S510, an abnormal behavior confirmation mechanism and baseline protection strategy are established. This involves setting dual judgment criteria to accurately identify and lock onto significant abnormal states of the subject. This step requires not only a drastic deviation in the overall behavior pattern of the current window or multiple consecutive windows (i.e., the standardized offset of the joint state representation vector exceeds a freezing threshold), but also that this drastic deviation must be reflected in a sufficient number of key behavioral features (i.e., the standardized offset of component state variables exceeding a preset threshold simultaneously exceeds their respective component thresholds). Once baseline freezing is triggered, the system will suspend updating the individual's historical baseline using current window data, effectively preventing abnormal behavior data from polluting long-term behavioral models and ensuring the stability and reference value of the baseline.

[0073] Step S520: When a subsequent preset number of consecutive windows meet the stabilization condition, recovery maintenance is performed on the individual historical baseline. Specifically, this includes: when the standardized offset of the joint state representation vector of a subsequent preset number of consecutive windows is lower than the recovery threshold, and the component state variables exceeding the preset number threshold re-enter their respective stable intervals, baseline recovery is triggered. If the recovery threshold is less than the freeze threshold, a hysteresis interval is formed.

[0074] In step S520, a hysteresis-based recovery mechanism is established to safely unlock the baseline lock after the subject's behavior returns to normal. This step sets stricter stabilization conditions, requiring that subsequent behavioral data from multiple consecutive windows return to the normal range. Specifically, the standardized offset of the joint state representation vector must be below the recovery threshold, and a sufficient number of key components must re-enter the stable interval before baseline recovery can be triggered. In particular, the recovery threshold is set to be lower than the freeze threshold. This design creates a hysteresis interval, effectively avoiding the jitter caused by frequent system freezes and recoveries due to small fluctuations in subject behavior around the threshold, thus improving the robustness and reliability of the entire behavior analysis system.

[0075] In some embodiments of this application, step S600, generating a structured state interpretation result based on the state comparison result, specifically includes the following steps.

[0076] Step S610: Generate a master state label based on the offset direction information and the dominant offset component information, and determine the trend state based on the continuous trend information, wherein the trend state includes continuous offset, continuous recovery, short-term fluctuation and stable maintenance.

[0077] In step S610, the abstract mathematical offset is transformed into human-readable behavioral semantic tags, completing the initial transformation from data to intelligence. This step generates a master state tag based on offset direction information and dominant offset component information, clearly indicating the specific abnormal characteristics currently exhibited by the subject, such as excessive speed or directional disorder. Simultaneously, based on continuous trend information, the trend state is determined, subdividing the behavioral pattern into categories such as continuous offset, continuous recovery, short-term fluctuations, and stable maintenance. This process not only describes the type of anomaly but also characterizes its development dynamics, enabling the system to distinguish between a one-off, accidental disturbance and an evolving, serious threat.

[0078] Step S620 involves feeding back the state interpretation results to the upstream anonymous long-cycle behavior chain reconstruction system. Specifically, this includes feeding back the state interpretation results, which contain information on the dominant offset component and risk processing participation identifiers, to the upstream system only when the trend state is a continuous shift and the state interpretation confidence is higher than the feedback threshold, thereby triggering chain-level risk scoring enhancement or manual review priority ranking.

[0079] In step S620, a precise feedback control mechanism is established to ensure that only high-value anomaly information is transmitted to the upper-level system, thereby optimizing computing resources and improving decision-making efficiency. This step sets strict screening conditions: only when the trend state shows a continuous shift and the state interpretation confidence is higher than the feedback threshold will the state interpretation result, which includes the dominant shift component information and the risk processing participation identifier, be fed back to the upstream anonymous long-cycle behavior chain reconstruction system.

[0080] This conditional feedback strategy effectively filters out low-value noise interference, preventing upstream systems from frequently adjusting due to minor fluctuations. Once feedback is triggered, the system will activate chain-level risk scoring enhancement or manual review and priority ranking, guiding system resources to focus on truly noteworthy high-risk targets, thus achieving an intelligent risk management closed loop.

[0081] In some embodiments of this application, the proposed method for screening historical stable windows, restoring baseline freezes, and providing chain-level risk feedback for anonymous long-cycle behavioral chains includes the following steps.

[0082] Step one involves obtaining anonymous chain window objects. Standardized data units, i.e., anonymous chain window objects, are obtained from the upstream anonymous long-cycle behavior chain reconstruction system. These objects are generated by long-cycle concatenation of the same subject in open street environments or multi-camera discrete coverage environments through pedestrian re-identification or cross-camera tracking technology.

[0083] To ensure the accuracy and privacy of subsequent analysis, the object is defined as a data structure containing a series of key attributes, including an anonymous chain identifier for uniquely identifying the subject, a window identifier and window time range for defining time slices, a region identifier describing the geographic location, and the core absolute ground physical trajectory sequence.

[0084] Furthermore, to quantify data reliability, the object also includes trajectory recovery quality information and chain confidence information. In a more preferred embodiment, to support internal system state transitions and tracking, the object is further extended to include window stability state information, baseline maintenance state information, and interruption recovery state information, thereby providing a complete data foundation for end-to-end risk detection and baseline maintenance.

[0085] Step two involves constructing component state variables, transforming the raw, low-level physical trajectory data into high-level, quantifiable behavioral characteristic indicators, thereby achieving a mathematical abstraction of the subject's motion pattern. Based on the absolute ground physical trajectory sequence within the current window, the system calculates and constructs at least three component state variables to describe the subject's behavioral characteristics from multiple dimensions.

[0086] In a preferred embodiment, the system first calculates the average displacement magnitude of adjacent trajectory points, using it as the average displacement state quantity to reflect the average range of activity of the subject; secondly, it calculates the average displacement per unit time, using it as the average velocity state quantity to characterize the speed of the subject's movement; thirdly, it calculates the variance of the change in the direction angle of the continuous trajectory, using it as the discrete direction change state quantity to measure the tortuosity or hesitant characteristics of the subject's movement trajectory; in addition, the system also counts the number of times the speed crosses a preset stop threshold, using it as a stop-start count state quantity to identify whether the subject has an intermittent stop-go pattern; finally, it calculates the periodic fluctuation amplitude of the velocity sequence or displacement sequence, using it as a periodic fluctuation state quantity to detect whether the subject has a regular reciprocating motion. These component state quantities together constitute a feature set describing the subject's behavior.

[0087] Step 3: Form a joint state representation vector and perform historical stability window filtering; the system concatenates multiple component state variables in a preset order to form a joint state representation vector. If the aforementioned five component state variables are used, the joint state representation vector can be expressed by the following formula: ;in, , , , and These correspond to the average displacement, average velocity, discrete direction change, stop-start switching count, and periodic fluctuation state quantities, respectively.

[0088] Building upon this foundation, this step further performs rigorous historical stable window screening on historical window objects corresponding to the same anonymous chain identifier. The screening process employs a multi-level filtering mechanism: first, low-quality windows with trajectory recovery quality below a first threshold or chain confidence below a second threshold are removed; then, the joint state representation vector standardized offset and component state quantity standardized offset of each historical window relative to the current individual historical baseline are calculated; finally, only those windows whose joint state representation vector standardized offset is below the first offset threshold, have at least P component state quantities within their respective stable intervals, and whose N consecutive windows satisfy the above conditions are retained as historical stable windows. This process ensures that the historical data used to establish the baseline is clean, reliable, and behaves normally.

[0089] Furthermore, based on historical stability windows, an individual historical baseline is established, and a mathematical center for the individual's normal behavioral pattern is established through statistical methods, i.e., a standardized reference benchmark is constructed. The system calculates the joint state representation vector of the historical stability window samples, and obtains its mean vector and historical fluctuation range vector, which serve as the baseline of the joint state representation vector; at the same time, the historical mean and historical fluctuation range are calculated for each component state quantity, which serve as the baseline of the component state quantity.

[0090] This process is equivalent to creating a behavioral profile for each anonymous subject, recording the average level and normal fluctuation boundaries of various motion indicators under stable conditions. At the same time, the system also records historical stable intervals, providing a basis for subsequent judgments on whether the behavior has returned to normal.

[0091] Step four involves real-time monitoring of the deviation between current behavior and historical norms, identifying potential risks by quantifying these differences. The system compares the component state variables and joint state representation vector corresponding to the current window with the individual's historical baseline to generate a state comparison result. Specific comparison metrics preferably include the standardized offset of the joint state representation vector, the standardized offset of the component state variables, offset direction information, and continuous trend information.

[0092] Among these, continuous trend information is particularly crucial. By analyzing the time series characteristics of the offset, it can effectively distinguish between different states such as continuous offset, continuous recovery, short-term fluctuations, and stable maintenance, thereby helping the system understand whether the current anomaly is a temporary noise interference or a behavioral pattern change with long-term impact.

[0093] Step 5: When the system detects that the standardized offset of the joint state representation vector of the current window or K consecutive windows is higher than the freezing threshold, and the standardized offset of at least P component state variables exceeds the corresponding component threshold, the system determines that the subject has entered a continuous abnormal state and then performs baseline freezing.

[0094] While the system continues to perform state comparisons and generate structured state interpretation results during the frozen state, the window data from the frozen phase will be excluded from individual historical baseline updates. This mechanism prevents the baseline from being erroneously skewed due to sudden changes in subject behavior (such as theft or long-term habit changes), ensuring the sensitivity of risk detection.

[0095] Furthermore, a dynamic adaptive mechanism is established to enable the system to follow the long-term benign evolution of the agent's behavior. When the standardized offset of the joint state representation vector for the subsequent M consecutive windows is lower than the recovery threshold, and at least P component state variables re-enter their respective stable intervals, while the trajectory recovery quality and chain confidence meet the update conditions, the system determines that the agent's behavior has returned to stability and then performs baseline recovery.

[0096] At this point, the restored stable window will be reintegrated into the individual's historical baseline update. This process ensures that the baseline model can both withstand the interference of short-term anomalies and adapt to the long-term, stable behavioral pattern migration of the subject, maintaining the timeliness of the detection model.

[0097] Furthermore, by setting rigorous mathematical constraints, the system's detection logic is optimized to prevent frequent jumps in the judgment result near the threshold. Regarding parameter settings, the following conditions are preferably met: P is not less than half the total number of component state variables participating in the comparison, rounded up, to ensure that a judgment is triggered only when most features are consistently abnormal; N is determined based on the window length and historical fluctuation level; K is greater than or equal to 2, and preferably K is less than or equal to N; M is greater than or equal to K.

[0098] Furthermore, by setting a first offset threshold lower than the freeze threshold and a recovery threshold lower than the freeze threshold, a hysteresis interval is created between the freeze and recovery decisions. This hysteresis design effectively prevents the system from repeatedly jumping between freeze and recovery states due to minor data fluctuations, thus improving the system's robustness and stability.

[0099] Preferably, the system parameters are specifically configured as follows: the minimum number of components P required to filter stable windows is equal to 3, the number of consecutive windows that meet the conditions N is equal to 3, the number of consecutive windows that trigger freezing K is equal to 3, and the number of consecutive windows that trigger recovery M is equal to 4. Simultaneously, the first offset threshold is set to 1.00, the freezing threshold to 1.80, the recovery threshold to 1.10, and the feedback threshold to 0.85.

[0100] Preferably, the individual historical baseline mean vector constructed based on historical stable window data is represented as follows: The corresponding historical fluctuation range vector is represented as Assume the currently detected window state vector is These initial data provide a concrete numerical basis for subsequent offset calculations and risk assessments.

[0101] Furthermore, based on the aforementioned current window state vector and historical baseline data, the system calculates the standardized offset values ​​for each component as 2.15, 3.30, 3.14, 2.33, and 2.13. To comprehensively assess the deviation of the overall behavior, if the standardized offset of the joint state representation vector is taken as the average of the absolute values ​​of the offsets of each component, then the joint offset value is calculated. Due to the calculated If the offset values ​​of the five component state variables exceed the preset freeze threshold of 1.80, and the offset values ​​of all five component state variables exceed their respective component thresholds, this indicates that the behavior pattern of the current window differs significantly from the historical norm. Therefore, if the system detects that three consecutive windows meet the above offset conditions, it will determine that the subject is in a continuous abnormal state, thereby triggering the baseline freeze mechanism and stopping the updating of the historical baseline using the current abnormal data.

[0102] During subsequent monitoring after baseline freezing, if the system detects that the standardized offset values ​​of the joint state representation vector for four consecutive windows are 1.05, 0.98, 0.90, and 0.86, respectively, these values ​​are all below the preset recovery threshold of 1.10. Simultaneously, in each window, at least three component state variables re-enter their respective stable regions, this indicates that the subject's behavioral pattern has returned to normal or a new stable state.

[0103] Based on this continuous regression trend, the system determines that the abnormal state has ended and then triggers the baseline recovery mechanism, which re-incorporates the recovered stable window data into the update calculation of the individual's historical baseline to ensure that the baseline model can dynamically adapt to changes in the subject's behavior.

[0104] Step Six: Generate a structured state interpretation result based on the state comparison results. This result preferably includes a main state label, trend state, state interpretation confidence level, dominant offset component information, and risk processing participation identifier. Regarding the feedback mechanism, the system will only feed back the structured state interpretation result to the upstream anonymous long-cycle behavior chain reconstruction system when the trend state exhibits a continuous offset and the state interpretation confidence level is higher than the feedback threshold. This triggers chain-level risk scoring enhancement, in-chain anomaly accumulation statistics, report generation triggering, or manual review priority ranking. Conversely, if the trend state only experiences short-term fluctuations, or the state interpretation confidence level is lower than the feedback threshold, the system will only record the interpretation result without triggering chain-level risk feedback, thereby effectively filtering out non-persistent noise interference and ensuring the accuracy of risk warnings.

[0105] In some embodiments of this application, reference is made to Figures 2 to 5 This demonstrates the complete process of the baseline detection and risk feedback method for anonymous behavioral chains provided in this application.

[0106] first, Figure 2 The diagram illustrates the process of obtaining an anonymous chain window object. The image shows multiple consecutive video frames, each featuring a cyclist marked with green skeleton lines and yellow keypoints, clearly outlining their posture. The core of this step is to extract motion data of the cyclist within a specific time window from the upstream anonymous long-period behavior chain reconstruction system, forming an anonymous chain window object. This object not only contains image information but also implicitly represents their continuous trajectory in time and space, providing the raw data foundation for subsequent analysis.

[0107] Secondly Figure 3This diagram illustrates the process of constructing component state variables. The left side of the image shows a cyclist labeled with skeletons, while the right side lists five component state variables extracted from the cyclist's trajectory: mean displacement, mean velocity, discrete direction change, stop-start switch count, and periodic fluctuation state variable. These components together constitute the joint state representation vector Z, whose mathematical expression is: This step transforms the original motion trajectory into quantifiable, multi-dimensional behavioral features, providing a mathematical foundation for subsequent baseline establishment and anomaly detection.

[0108] Furthermore, Figure 4 The diagram illustrates the process of establishing an individual historical baseline and detecting anomalies, where the horizontal axis represents time and the vertical axis represents the state variables of the i-th individual. The scattered black dots in the diagram represent historical observation samples of that individual at different time points; these samples collectively constitute the data foundation for baseline modeling. Based on this historical data, the system fits a smooth regression baseline. This is used to characterize the individual's normal behavioral trend over time. The area enclosed by dashed lines on both sides of the baseline is the stable fluctuation zone, defining the permissible deviation of the individual's behavior within the normal range. At the current time t, the current observation value... This is represented as a black dot deviating from the baseline, and its perpendicular distance from the baseline is the error term. Satisfying the formula When this error exceeds the threshold of the stable fluctuation band, it is judged as abnormal behavior. The entire process reflects the modeling logic of learning normal patterns from historical data and using this as a basis for real-time anomaly detection.

[0109] at last, Figure 5 The process of baseline freezing and recovery is illustrated, and the figure shows the absolute value of error over time. The system tracks the behavior of an individual's regression baseline. When the curve remains above the "freeze threshold," the system identifies it as a "continuous deviation freeze zone" and enters a "frozen" state. During this period, the abnormality persists without contaminating the individual's regression baseline. When the curve falls back and remains below the "recovery threshold," the system identifies it as a "continuous stabilization recovery zone" and enters a "recovery" state. This step, by setting up freeze and recovery mechanisms, ensures that the baseline is not incorrectly updated when the subject's behavior is persistently abnormal, and that it can dynamically adapt to the new normal after the behavior returns to normal, thus guaranteeing the accuracy and robustness of risk detection. Finally, the system performs subsequent processing based on the detection results, such as "risk score enhancement / abnormality accumulation / review and ranking."

[0110] In some embodiments of this application, to verify the robustness of the baseline detection and risk feedback method, 1000 consecutive window objects under the same anonymous chain identifier were selected for testing. Among them, 800 windows were stable or short-term fluctuating windows (normal behavior), and 200 windows were continuously offset windows (simulating abnormal behavior). The specific experimental setup is as follows: (1) Control group setup: The ordinary moving average baseline method was used.

[0111] (2) Experimental group setup: The historical stability window screening, baseline freeze recovery and structured risk feedback methods described in this application were adopted.

[0112] Based on the above experimental setup, the experimental results are as follows: (1) Comparison of false alarm rates: The control group generated 96 false alarms in 800 non-abnormal windows, with a false alarm rate of 12.0%; the experimental group generated 37 false alarms, with a false alarm rate of 4.6%.

[0113] (2) Comparison of baseline contamination after the end of the continuous offset phase: The baseline offset of the control group was 1.42, indicating that the baseline had been severely skewed by abnormal data; the baseline offset of the experimental group was 0.38, indicating that the freezing mechanism effectively suppressed the contamination of the baseline by abnormal data.

[0114] Therefore, this application can effectively suppress the contamination of an individual's historical baseline by abnormal data and significantly improve the accuracy of risk feedback.

[0115] Secondly, refer to Figure 6 This application provides a baseline detection and risk feedback system for anonymous behavior chains, including an object acquisition module, a state construction module, a baseline maintenance module, and a risk feedback module.

[0116] In some embodiments of this application, the object acquisition module is configured to: acquire the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system, wherein the anonymous chain window object includes at least the anonymous chain identifier, the window identifier, the window time range, and the absolute ground physical trajectory sequence.

[0117] Specifically, the object acquisition module, serving as the entry point for the entire system, establishes a standardized data interface with the upstream anonymous long-cycle behavior chain reconstruction system. This module is responsible for receiving anonymous chain window objects output in open street environments or multi-camera discrete coverage environments. These objects have already been concatenated using pedestrian re-identification or cross-camera tracking technologies. By acquiring basic information including anonymous chain identifiers, window identifiers, window time ranges, and absolute ground physical trajectory sequences, this module ensures that the data units for subsequent analysis possess spatiotemporal continuity and privacy security, laying a solid data foundation for end-to-end risk detection.

[0118] In some embodiments of this application, the state construction module is configured to: construct multiple component state quantities based on the absolute ground physical trajectory sequence in the anonymous chain window object, and concatenate the multiple component state quantities to form a joint state representation vector.

[0119] Specifically, the state construction module transforms the raw trajectory coordinate data into a quantifiable, multi-dimensional feature space representation. Based on the acquired absolute ground physical trajectory sequence, this module deeply mines and constructs multiple component state quantities, including mean displacement, mean velocity, dispersion of direction changes, stop-and-go switching counts, and periodic fluctuations. Subsequently, it concatenates these components reflecting different motion characteristics of the subject to form a joint state representation vector, achieving a comprehensive abstraction of the subject's behavioral patterns within the window period. This high-dimensional vectorization not only enhances the distinguishability of features but also provides a standardized mathematical model for subsequent accurate comparison with individual historical baselines.

[0120] In some embodiments of this application, the baseline maintenance module is configured to: for historical window objects corresponding to the same anonymous chain identifier, filter historical stable windows based on trajectory recovery quality, chain confidence, standardized offset of joint state representation vector, and the number of consecutive stable windows, and establish individual historical baselines based on the historical stable windows; compare the component state variables and joint state representation vectors corresponding to the current window with the individual historical baselines to form a state comparison result; when the current window or a consecutive preset number of windows meet the continuous offset condition, perform freeze maintenance on the individual historical baselines. When a subsequent consecutive preset number of windows meet the stabilization condition, perform recovery maintenance on the individual historical baselines.

[0121] Specifically, the baseline maintenance module first selects high-quality historical stability windows based on multiple stringent conditions such as trajectory recovery quality and chain confidence, thereby constructing an individual historical baseline that truly reflects the normal habits of the subject. Secondly, it compares the state vector of the current window with the historical baseline, calculating a standardized offset to form the state comparison result. More importantly, this module introduces a dynamic freeze and restore mechanism: when a continuous behavioral shift is detected, baseline updates are automatically frozen to prevent anomalous data contamination; and maintenance is restored after the behavior stabilizes, ensuring that the baseline maintains stability while adapting to long-term behavioral evolution.

[0122] In some embodiments of this application, the risk feedback module is configured to: generate a structured state interpretation result based on the state comparison result, and when the trend state is continuously shifting and the confidence level of the state interpretation is higher than the feedback threshold, feed the state interpretation result back to the upstream anonymous long-cycle behavior chain reconstruction system for chain-level risk processing.

[0123] Specifically, the risk feedback module generates a structured interpretation result containing a main state label and a trend state based on the state comparison results, and sets strict trigger thresholds. Only when the trend state shows a continuous shift and the confidence level of the state interpretation is higher than a preset threshold will the result containing the dominant shift component information and risk processing participation identifier be fed back to the upstream system. This precise feedback mechanism can effectively filter noise interference and directly trigger the upstream system's chain-level risk scoring enhancement or manual review priority ranking, thereby significantly improving the overall security system's proactive defense capabilities and response efficiency.

[0124] Furthermore, the structured state interpretation results output by the risk feedback module specifically include risk scoring enhancement instructions, manual review priority identifiers, and abnormal behavior feature template update packages. This feedback result is transmitted to the upper-layer behavior chain reconstruction architecture layer (i.e., the upstream anonymous long-cycle behavior chain reconstruction system) to dynamically adjust the weight parameters in the association cost function, thereby achieving adaptive tracking optimization based on risk state.

[0125] Furthermore, embodiments of this application provide an electronic device, including: one or more processors; and a memory for storing one or more programs. When the one or more programs are executed by the one or more processors, the processors perform the baseline detection and risk feedback method for anonymous behavior chains as described above.

[0126] Furthermore, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the baseline detection and risk feedback method for anonymous behavioral chains as described above.

[0127] In summary, the baseline detection and risk feedback method and related equipment for anonymous behavior chains provided in this application have the following technical effects.

[0128] This technical solution effectively improves the computability and interpretability of window-level state comparison by introducing anonymous chain window objects as restricted input boundaries and combining them with at least three component state variables and a joint state representation vector. The method utilizes multi-dimensional conditions such as trajectory recovery quality, chain confidence, and standardized offset to screen historically stable windows, thereby establishing a representative individual historical baseline. This solves the problem of baseline inaccuracy caused by directly using low-quality historical data in existing technologies, thus improving the accuracy of individual behavior pattern modeling.

[0129] By implementing a baseline freezing and recovery mechanism with continuous window linkage, this application can pause baseline updates when a persistent shift is detected, effectively preventing abnormal behavioral data from contaminating individual historical baselines. Updates are then dynamically resumed after the behavior stabilizes, enhancing the model's adaptability to long-term behavioral changes. Simultaneously, by setting a joint trigger threshold for trend status and explanatory confidence, chain-level risk feedback is executed only when a persistent shift is met and the confidence level is sufficient, significantly reducing the false alarm rate and improving the accuracy of risk warnings and the processing efficiency of the upstream system.

[0130] 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.

[0131] 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.

[0132] 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.

[0133] 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.

[0134] 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.

[0135] 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.

[0136] 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.

[0137] 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.

[0138] 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.

[0139] 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 baseline detection and risk feedback method for anonymous behavioral chains, characterized in that, The method includes the following steps: Obtain the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system. The anonymous chain window object includes at least the anonymous chain identifier, window identifier, window time range, and absolute ground physical trajectory sequence. Multiple component state quantities are constructed based on the absolute ground physical trajectory sequence in the anonymous chain window object, and the multiple component state quantities are concatenated to form a joint state representation vector. For historical window objects corresponding to the same anonymous chain identifier, historical stable windows are selected based on trajectory recovery quality, chain confidence, standardized offset of joint state representation vector, and number of continuous stable windows, and individual historical baselines are established based on the historical stable windows. The component state quantity and joint state representation vector corresponding to the current window are compared with the individual historical baseline to form a state comparison result; When the current window or a consecutive preset number of windows meet the continuous offset condition, the individual historical baseline is frozen and maintained; when the subsequent consecutive preset number of windows meet the stabilization condition, the individual historical baseline is restored and maintained. A structured state interpretation result is generated based on the state comparison result. When the trend state is continuously shifting and the confidence level of the state interpretation is higher than the feedback threshold, the state interpretation result is fed back to the upstream anonymous long-cycle behavior chain reconstruction system for chain-level risk handling.

2. The baseline detection and risk feedback method for anonymous behavioral chains according to claim 1, characterized in that, The process of obtaining the anonymous chain window object output by the upstream anonymous long-cycle behavior chain reconstruction system specifically includes: In open street blocks or pedestrian environments with multiple cameras discretely covered, receive the anonymous chain window object output by the upstream anonymous long-period behavior chain reconstruction system, which has been connected by pedestrian re-identification or cross-camera tracking technology. The anonymous chain window object also includes trajectory recovery quality information and chain confidence information.

3. The baseline detection and risk feedback method for anonymous behavioral chains according to claim 1, characterized in that, The component state quantities specifically include: The average displacement state quantity is used to characterize the average movement distance within the current window, and is determined based on the average displacement magnitude of adjacent trajectory points within the current window. The average velocity state quantity is used to characterize the average movement speed within the current window, and is determined based on the average displacement per unit time within the current window. The direction change discrete state quantity is used to characterize the stability of the motion direction within the current window. It is determined based on the variance, standard deviation, or quantile dispersion of the continuous trajectory direction angle change within the current window. The stop-go switching count is used to characterize the frequency of motion pauses within the current window, and is determined based on the number of times the speed crosses the stop-go threshold within the current window. Periodic fluctuation state quantities are used to characterize the motion rhythm within the current window. They are determined based on the periodic amplitude, dominant frequency energy, or periodic fluctuation range of the velocity sequence, displacement sequence, or gait rhythm sequence within the current window.

4. The baseline detection and risk feedback method for anonymous behavioral chains according to claim 1, characterized in that, The process of selecting historical stable windows for the same anonymous chain identifier based on trajectory recovery quality, chain confidence, standardized offset of the joint state representation vector, and the number of continuous stable windows specifically includes: Historical stable windows are defined as those whose trajectory recovery quality is higher than a first threshold, whose chain confidence is higher than a second threshold, whose joint state representation vector standardized offset is lower than a first offset threshold, and whose historical windows satisfy the stability condition for a consecutive preset number of times. The mean vector of the joint state representation vector of the selected historical stable window samples is calculated and used as the baseline of the joint state representation vector. Calculate the historical fluctuation range vector for the historical stable window sample, which serves as the fluctuation boundary of the individual historical baseline; The historical fluctuation range vector is composed of a standard deviation vector, a quantile interval vector, or a robust fluctuation range vector.

5. The baseline detection and risk feedback method for anonymous behavioral chains according to claim 1, characterized in that, The step of comparing the component state quantities and joint state representation vector corresponding to the current window with the individual historical baseline to form a state comparison result specifically includes: Calculate the standardized offset of the current window joint state representation vector relative to the individual historical baseline, and the component standardized offset of each component state quantity relative to the corresponding component baseline. The state comparison result includes at least the standardized offset, component standardized offset, offset direction information, and continuous trend information.

6. The baseline detection and risk feedback method for anonymous behavioral chains according to claim 1, characterized in that, When the current window or a consecutive preset number of windows meet the continuous offset condition, the individual historical baseline is frozen and maintained, specifically including: When the standardized offset of the joint state representation vector of the current window or a consecutive preset number of windows is higher than the freezing threshold, and the standardized offset of the component state quantity exceeding the preset number threshold exceeds the corresponding component threshold, baseline freezing is triggered; in the frozen state, the current window does not participate in the update of the individual historical baseline. When a subsequent preset number of windows meet the stabilization condition, the individual historical baseline is restored and maintained, specifically including: When the standardized offset of the joint state representation vector of a subsequent preset number of windows is lower than the recovery threshold, and the component state quantities exceeding the preset number threshold re-enter their respective stable intervals, baseline recovery is triggered; the recovery threshold is less than the freeze threshold, forming a hysteresis interval.

7. The baseline detection and risk feedback method for anonymous behavioral chains according to claim 1, characterized in that, The generation of structured state interpretation results based on the state comparison results specifically includes: The main state label is generated based on the offset direction information and the dominant offset component information, and the trend state is determined based on the continuous trend information. The trend state includes continuous offset, continuous recovery, short-term fluctuation and stable maintenance. The step of feeding back the state interpretation results to the upstream anonymous long-cycle behavior chain reconstruction system specifically includes: Only when the trend state is a continuous shift and the state interpretation confidence is higher than the feedback threshold, the state interpretation result, which includes information on the dominant shift component and the risk processing participation identifier, will be fed back to the upstream system, triggering chain-level risk scoring enhancement or manual review priority ranking.

8. A baseline detection and risk feedback system for anonymous behavioral chains, characterized in that, Steps for performing the baseline detection and risk feedback method for anonymous behavioral chains as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the processors perform the baseline detection and risk feedback method for anonymous behavior chains as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the baseline detection and risk feedback method for anonymous behavioral chains as described in any one of claims 1 to 7.

Images