Abnormal behavior real-time identification and intelligent early warning system based on video stream structured analysis

Abstract

The invention relates to the technical field of public safety monitoring, and discloses an abnormal behavior real-time identification and intelligent early warning system based on video stream structured analysis. The objective of the invention is to solve the problems that hidden threat perception of an existing monitoring system under the condition of clothes shielding is lacked, and missing report and false report pain points caused by incapability of distinguishing physiologic pathological conditions and subjective abnormal behaviors. Through the explicit-implicit double-track mechanism and negative weight correction of the pathological manifold space, immediate blocking of explicit violence and perspective perception of implicit threats are achieved. The core of the method is that microscopic posture features and multi-dimensional risk vectors are fused, and a closed-loop decision system with humanitarian care is constructed. According to the method, the recognition precision of the hidden high-risk target is remarkably improved, and the anti-interference capability and the social suitability of the system are greatly enhanced.

Description

Technical Field

[0001] This invention relates to the field of public safety monitoring technology, and more specifically, to a real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis. Background Technology

[0002] As a core infrastructure of modern urban security and prevention systems, public safety monitoring systems play an irreplaceable role in maintaining social stability and protecting the personal and property safety of citizens. With the deepening of smart security construction, large-scale video surveillance networks have widely covered airports, customs ports, transportation hubs, and various densely populated areas. These systems, through 24 / 7, all-round video collection, continuously generate massive amounts of visualized data resources, playing a crucial role in deterring illegal and criminal activities, assisting urban emergency management, and facilitating post-crime evidence tracing. They serve as the digital visual hub for building a three-dimensional social security and prevention system.

[0003] However, despite significant advancements in facial recognition and basic behavior detection, current intelligent video analytics technologies still face substantial technical challenges in addressing complex and concealed security threats. First, existing behavior recognition algorithms primarily rely on capturing the range of macroscopic limb movements, lacking penetrating perception capabilities for concealed threats hidden under clothing (such as concealed knives or firearms). This results in security systems often only responding passively after a violent incident, failing to identify high-risk targets carrying dangerous items and exhibiting nervous behavior beforehand. Second, existing systems lack the ability to finely differentiate physiological pathological characteristics in their anomaly detection logic. They are prone to misidentifying atypical gaits of Parkinson's patients, stroke survivors, or individuals with physical disabilities as suspicious behavior. This not only leads to frequent false alarms and a significant waste of security resources but also renders the technical solutions lacking necessary humanitarian considerations and social adaptability in practical applications. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis, which solves the problems of existing monitoring systems that lack perception of hidden threats hidden under clothing and cannot distinguish between physiological pathological conditions and subjective abnormal behavior, resulting in missed and false alarms.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a real-time abnormal behavior identification and intelligent early warning system based on video stream structured analysis, which consists of a set of highly coordinated functional modules, specifically including: Skeleton attachment module: Locates the human body region in the image frame of the synchronous video data stream, extracts the feature response extreme points in the region, connects the feature response extreme points into a skeleton topology according to human kinematic constraints, and converts the pixel coordinates into normalized coordinates; Cross-view labeling module: It integrates human appearance texture and skeletal motion cycle to generate feature vectors, and based on the matching results of the feature vectors and the global trajectory library, it unifies the identity of the same target under different viewpoints; Explicit risk assessment module: calculates the kinetic energy entropy evolution rate of the interaction region of the skeleton limb nodes, and generates a physical danger blocking signal when the kinetic energy entropy evolution rate exceeds the preset safety boundary; Potential risk assessment module: integrates the risk value of concealed foreign objects, the proportion of postural stress risk, and the trajectory discrete entropy value to generate the initial hidden risk coefficient corresponding to the identity identifier; Overall risk assessment module: Upon receiving a physical hazard blocking signal, it directly locks the alarm status; it comprehensively assesses the latent hazard coefficient and the object's hazard level after correction by the negative weight exemption coefficient, and when the total risk value continues to exceed the threshold, it marks the corresponding identity as an abnormal risk target and outputs a risk warning signal.

[0007] As a preferred embodiment of the real-time abnormal behavior identification and intelligent early warning system based on video stream structured analysis described in this invention, the system further includes the following modules: Data alignment module: It uses a built-in clock to uniformly calibrate multiple monitoring devices and align the timestamps of multiple video streams to form a synchronized video data stream; Hidden risk calculation module: Extracts the deformation vector of the human clothing surface, and based on the phase and amplitude difference between the deformation vector and the skeletal movement, identifies non-natural damping areas and generates hidden foreign object risk values; Postural rigidity calculation module: calculates the symmetry error of the movement trajectory of the left and right limbs, and extracts high-frequency micro-tremor features from the vertical movement sequence of the shoulder and neck nodes to generate the proportion of postural stress risk; Trajectory wandering degree calculation module: maps the movement trajectory of the identity identifier into a discrete grid sequence, and calculates the discrete entropy value of the trajectory based on the spatiotemporal repeated access frequency; Pathological behavior determination module: Calculates the similarity distance between the current spatiotemporal sequence of the skeleton and the preset pathological gait model, and generates a negative weight exemption coefficient based on the similarity distance.

[0008] As a preferred embodiment of the real-time abnormal behavior recognition and intelligent early warning system based on video stream structured analysis described in this invention, the specific process of the skeleton attachment module locating and extracting the skeleton topology includes: Perform multi-scale feature extraction and convolution operations on the image frame to identify regions with human appearance texture features and generate human body bounding boxes; Calculate the feature response intensity of pixel data within the human body bounding box to key anatomical parts of the human body, generate a probability response heatmap including the top of the head, neck, shoulder, elbow, wrist, hip, knee and ankle, and select the local response extreme points in the heatmap as key skeletal points. A local Cartesian coordinate system is constructed with the midpoint of the line connecting the left and right hip joints as the origin and the pixel height of the human body bounding box as the unit length. The absolute pixel coordinates of the key bone points are projected into the local Cartesian coordinate system to generate the skeleton topology that eliminates the difference between shooting distance and viewpoint scale.

[0009] As a preferred embodiment of the real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis described in this invention, the specific process of unifying the identity identifier in the cross-view domain marking module includes: Extract the color distribution histogram and texture gradient features of the human body region, and calculate the gait interaction frequency and stride features of the bipedal nodes in the skeleton topology. Concatenate and stitch the color distribution histogram, texture gradient features and gait features into a multidimensional identity description vector. Calculate the cosine similarity between the current multidimensional identity description vector and the centroids of each stored feature vector set in the global trajectory database; If the calculated highest cosine similarity exceeds the preset identity association threshold, the current multidimensional identity description vector is assigned to the corresponding feature vector set, and the centroid of the feature vector set is updated as the identity matching template for the corresponding target in the next moment. If the calculated highest cosine similarity is lower than the identity association threshold, the current target is determined to be a newly emerging target and a new globally unique identity is assigned. A new feature vector set is initialized in the global trajectory database.

[0010] As a preferred embodiment of the real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis described in this invention, the specific process by which the explicit risk assessment module generates a physical danger blocking signal includes: Real-time monitoring of wrist and ankle nodes of the human skeleton within the field of view; when the spatial Euclidean distance between skeleton nodes is less than a preset contact threshold, the corresponding spatial range is locked as the interactive area. Calculate the velocity vectors of all nodes within the interaction region, calculate the regional kinetic energy entropy based on the amplitude distribution variance and direction dispersion of the velocity vectors, and obtain the kinetic energy entropy evolution rate by differentiating the regional kinetic energy entropy over a continuous time series. The kinetic energy entropy evolution rate is compared with a preset biomechanical violence limit threshold. If the kinetic energy entropy evolution rate shows a positive surge and the peak value exceeds the violence limit threshold, it is determined that a physical dangerous behavior has occurred, and the physical danger blocking signal is generated. Furthermore, if the kinetic entropy evolution rate is detected to exceed the violent limit threshold, but no interactive area satisfying the contact threshold is formed between the skeleton nodes, the current behavior is identified as a normal abnormal behavior.

[0011] As a preferred embodiment of the real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis described in this invention, the specific process by which the hidden risk calculation module generates the hidden foreign object risk value includes: Calculate the instantaneous motion vector field of clothing pixels within the human torso region, and calculate the translational velocity vector of the corresponding bone nodes; Subtract the translational velocity vector from the instantaneous motion vector field to separate the relative swaying component of the clothing relative to the human torso; Calculate the amplitude of the relative sway component and the phase lag angle of the corresponding component relative to the human gait cycle; If the relative oscillation component amplitude in a local area is detected to be close to zero or the phase lag angle is found to be unnaturally locked, it is determined that there is hard object traction or gravity damping in the corresponding area, and the attenuation degree of the amplitude is normalized and mapped to a hidden foreign object risk value. Furthermore, the specific process by which the postural rigidity calculation module generates the proportion of postural stress risk includes: Using the line connecting the center of the human torso as the axis of symmetry, the spatiotemporal motion trajectory of the skeletal nodes of the left limb is mirrored and mapped to the coordinate system of the right limb. The root mean square error of the trajectory overlap between the left and right limbs in a unit gait cycle is calculated, and the root mean square error is normalized to the limb asymmetry coefficient. Vertical displacement data of the shoulder and neck nodes within a continuous time window are extracted, and the vertical displacement data are transformed from the time domain to the frequency domain using Fourier transform. The proportion of spectral energy in the 8Hz to 12Hz physiological tremor frequency band in the total spectral energy is calculated as the high-frequency microtremor coefficient. The limb asymmetry coefficient and the high-frequency microtremor coefficient are weighted and summed to obtain the proportion of postural stress risk that characterizes the target's psychological tension or physiological rigidity.

[0012] As a preferred embodiment of the real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis described in this invention, the specific process of the trajectory wandering degree calculation module calculating the trajectory discrete entropy value includes: The monitoring field of view is divided into a two-dimensional discrete grid of fixed size, and the coordinate points of the motion trajectory of the identity identifier within a continuous time window are mapped to a grid index sequence. By statistically analyzing the frequency of recurrence of each grid index in the sequence and the dwell time of the target in each grid, a spatial distribution probability model of the target's movement trajectory is constructed. The spatial Shannon entropy of the trajectory is calculated based on the spatial distribution probability model. At the same time, the cumulative angle of the direction change of the target motion vector is calculated. The spatial Shannon entropy and the cumulative angle of direction change are normalized and fused to generate a discrete entropy value of the trajectory that reflects the degree of disorder of the target's movement route and the degree of wandering in the region.

[0013] As a preferred embodiment of the real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis described in this invention, the specific process by which the potential risk assessment module generates the initial implicit risk coefficient includes: The hidden foreign object risk value, the proportion of body stress risk, and the trajectory discrete entropy value are respectively subjected to interval mapping normalization processing. The normalized values ​​are used as independent orthogonal components to construct a three-dimensional feature vector representing the hidden state of the target. The Euclidean modulus of the three-dimensional feature vector is calculated and used as the initial hidden risk coefficient to quantify the potential threat level of the corresponding identity identifier.

[0014] As a preferred embodiment of the abnormal behavior real-time identification and intelligent early warning system based on video stream structured analysis described in this invention, the specific process of the pathological behavior determination module generating the negative weight exemption coefficient includes: Skeletal motion data containing various typical pathological gaits are pre-collected, and a pathological gait kinematic manifold space that maps to a low-dimensional feature distribution is constructed. Calculate the inter-frame Euclidean distance matrix between the real-time spatiotemporal sequence of the current target's skeleton and the standard template in the pathological gait kinematic manifold space, and find a time warping path with the minimum cumulative distance in the matrix to perform temporal alignment and eliminate motion rate differences; Calculate the geodesic distance between the projection point of the aligned spatiotemporal sequence of bones in the manifold space and the center of the manifold. The geodesic distance is mapped to a negative weight exemption coefficient based on a preset exponential decay function, wherein the smaller the geodesic distance, the larger the absolute value of the generated negative weight exemption coefficient.

[0015] As a preferred embodiment of the real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis described in this invention, the specific process of the overall risk assessment module performing comprehensive assessment and early warning includes: Establish a monitoring channel for the physical hazard blocking signal. Upon detecting the physical hazard blocking signal, lock the risk status of the current identity to the highest level and trigger an alarm. If no physical hazard blocking signal is detected, obtain the object detection result associated with the current identity identifier, and query the preset dangerous item level table to determine the corresponding object hazard weight; The initial latent hazard coefficient and the negative weight exemption coefficient are algebraically superimposed and corrected, and the corrected result is added to the object hazard weight to obtain the real-time total risk value; A fixed-length sliding time window queue is constructed to store the real-time total risk value of historical frames. If the real-time total risk value of all data frames in the corresponding queue is higher than the preset composite warning threshold, an anomaly is confirmed, the identity identifier is marked as an abnormal risk target, and a warning signal containing the anomaly type and evidence fragments is output.

[0016] The beneficial effects of this invention are as follows: This invention achieves rapid alerts after violent incidents by monitoring the kinetic entropy evolution rate of the interaction areas of limb nodes; it overcomes the technical blind spot of traditional vision in identifying concealed foreign objects by utilizing the phase difference and amplitude attenuation analysis of clothing optical flow and skeletal movement; it keenly captures microscopic postural abnormalities caused by psychological stress by constructing a limb mirror-symmetric model and extracting 8-12Hz physiological micro-tremor features of the shoulder and neck; and it eliminates gait interference caused by physiological diseases by combining a negative weight exemption coefficient generated based on manifold learning and dynamic time warping technology; simultaneously, it maintains identity consistency across all scenarios by integrating appearance and gait features, and filters video noise using sliding time windows and multi-factor algebraic superposition logic. The synergistic effect of the above technical means jointly constructs a real-time abnormal behavior identification and intelligent early warning system that possesses both the ability to hard-block overt violent behavior and the ability to penetrate disguises to identify concealed suspicious persons, while also taking into account humanitarian concerns and extremely robustness. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a framework diagram of a real-time abnormal behavior identification and intelligent early warning system based on structured video stream analysis.

[0019] Figure 2 A flowchart for cross-viewpoint identity tagging and tracking.

[0020] Figure 3 Flowchart for calculating explicit and implicit dual-track risks.

[0021] Figure 4 This is a flowchart of pathological exemptions and overall decision-making. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.

[0025] Example 1 Reference Figures 1-4 This is one embodiment of the present invention, which provides a real-time identification and intelligent early warning system for abnormal behavior based on video stream structured analysis. This system is typically deployed on a high-performance computing server.

[0026] The system's software architecture consists of a set of highly collaborative functional modules, specifically including: Data alignment module: It uses a built-in clock to uniformly calibrate multiple monitoring devices and align the timestamps of multiple video streams to form a synchronized video data stream; Skeleton attachment module: Locates the human body region in the image frame of the synchronized video data stream, extracts the feature response extreme points within the region, connects the feature response extreme points into a skeleton topology according to human kinematic constraints, and converts the pixel coordinates into normalized coordinates; The skeleton attachment module locates and extracts the skeleton topology, performs multi-scale feature extraction and convolution operations on the image frame, identifies regions with human appearance texture features and generates human bounding boxes. Calculate the feature response intensity of pixel data within the human body bounding box to key anatomical parts of the human body, generate a probability response heatmap including the top of the head, neck, shoulder, elbow, wrist, hip, knee and ankle, and select the local response extreme points in the heatmap as key skeletal points. Construct a local Cartesian coordinate system with the midpoint of the line connecting the left and right hip joints as the origin and the pixel height of the human body bounding box as the unit length. Project the absolute pixel coordinates of key bone points into the local Cartesian coordinate system to generate a skeleton topology that eliminates the difference between shooting distance and viewpoint scale. Cross-view labeling module: It integrates human appearance texture and skeletal motion cycle to generate feature vectors. Based on the matching results of feature vectors and global trajectory library, it unifies the identity of the same target under different viewpoints. The specific process of unifying identity identification in the cross-view domain labeling module is to extract the color distribution histogram and texture gradient features of the human body region, and at the same time calculate the gait interaction frequency and stride features of the bipedal nodes in the skeleton topology. The color distribution histogram, texture gradient features and gait features are concatenated and spliced ​​into a multi-dimensional identity description vector. Calculate the cosine similarity between the current multidimensional identity description vector and the centroids of each stored feature vector set in the global trajectory database; If the calculated highest cosine similarity exceeds the preset identity association threshold, the current multidimensional identity description vector is assigned to the corresponding feature vector set, and the centroid of the feature vector set is updated as the identity matching template for the corresponding target in the next moment. If the calculated highest cosine similarity is lower than the identity association threshold, the current target is determined to be a newly emerging target and a new globally unique identity is assigned. A new feature vector set is initialized in the global trajectory database.

[0027] The system first connects to video streams from multiple network cameras covering the monitored area, constructing a unified spatiotemporal processing framework. To eliminate time errors caused by hardware clock drift between different front-end devices, the system uses a built-in NTP service module to send time synchronization commands to all front-end acquisition devices at a preset frequency (e.g., once every 60 seconds), ensuring the accuracy of timestamps for all video frames.

[0028] A timeline jitter buffer with a fixed length (e.g., 200 milliseconds) is established at the data entry point on the server side. The timestamp information in the header of each video data packet is parsed, and this buffer is used to reorder and align video streams from different sources. Expired frames with delays exceeding the buffer length are discarded, outputting a strictly synchronized, multi-view video data stream. This establishes a unified time reference for subsequent cross-camera target association and collaborative analysis.

[0029] The image processing unit performs multi-scale feature extraction and convolution operations on image frames. By constructing an image feature pyramid, regions of interest with human appearance texture features (such as upright shape and head and shoulder contours) are searched at different resolution levels, and bounding boxes surrounding human targets are generated.

[0030] By calculating the feature response intensity of pixel data within the bounding box to key anatomical locations, a set of probabilistic response heatmaps is generated. Each highlighted area in the heatmap represents the probability of the presence of a key location (such as the left elbow or right knee). By identifying local response extrema in the heatmaps, the pixel coordinates of key skeletal points such as the top of the head, neck, shoulder, elbow, wrist, hip, knee, and ankle are located. These points are then connected to form a complete skeletal topology according to the limb connection constraints of human biomechanics.

[0031] To eliminate perspective distortion caused by camera installation height, shooting angle, and target distance, coordinate normalization is also required. A local Cartesian coordinate system is constructed with the midpoint of the line connecting the left and right hip joints as the origin and the pixel height of the human body bounding box as the unit length. The absolute pixel coordinates of all extracted key skeletal points are projected into this local coordinate system, generating a set of relative normalized coordinates. At this point, regardless of whether the target is close at hand or far away in the frame, its skeletal topology has numerical scale consistency, providing standardized geometric input for subsequent motion analysis.

[0032] A multimodal feature fusion strategy is employed to generate an identity description vector. On one hand, the distribution histogram of the hue-saturation-brightness color space, which describes the target's appearance attributes, and local binary pattern texture gradient features are extracted as static visual features of the human body region. On the other hand, time-series data of the skeleton topology are used to calculate the gait interaction frequency (step frequency) and stride length features of bipedal nodes, representing the target's motion attributes. These two distinct feature sets are then concatenated and stitched together to form a high-dimensional, multi-dimensional identity description vector.

[0033] Subsequently, the generated multidimensional identity description vector is matched with a global trajectory library that stores the centroids of the feature vector sets of all active targets in the current scene (i.e., the average value of the target's features over a period of time). The cosine similarity between the current vector and each centroid vector is calculated, and the following state transition is performed based on the similarity results: If the calculated highest cosine similarity exceeds a preset identity association threshold (e.g., 0.85), the current target is determined to be an "old target" already existing in the database. In this case, not only is the current vector assigned to the target's feature set, but the centroid of the feature vector set is also updated using a weighted average algorithm. This dynamic update mechanism allows the system to adapt to feature drift caused by changes in lighting or pose during target movement, maintaining the freshness of the identity template at all times.

[0034] If the calculated highest cosine similarity is lower than the identity association threshold, the current target is determined to be a newly appeared target. Then, a new, globally unique identity is assigned to it, and a new feature vector set is initialized in the global trajectory database to formally start cross-camera tracking and recording of the target.

[0035] Explicit risk assessment module: Calculates the kinetic energy entropy evolution rate of the interaction area of ​​the skeleton limb nodes, and generates a physical danger blocking signal when the kinetic energy entropy evolution rate exceeds the preset safety boundary; The explicit risk assessment module generates physical hazard blocking signals in real time by monitoring the wrist and ankle nodes of the human skeleton within the field of view. When the spatial Euclidean distance between the skeleton nodes is less than the preset contact threshold, the corresponding spatial range is locked as the interaction area. Calculate the velocity vectors of all nodes within the interaction region, calculate the regional kinetic energy entropy based on the variance of the magnitude distribution and the directional dispersion of the velocity vectors, and obtain the kinetic energy entropy evolution rate by taking the derivative of the regional kinetic energy entropy over a continuous time series. The kinetic energy entropy evolution rate is compared with a preset biomechanical violence limit threshold. If the kinetic energy entropy evolution rate shows a positive surge and the peak value exceeds the violence limit threshold, it is determined that a physical dangerous behavior has occurred, and a physical danger blocking signal is generated. In addition, if the kinetic energy entropy evolution rate is detected to exceed the violent limit threshold, but no interaction area that meets the contact threshold is formed between the skeleton nodes, the current behavior is identified as normal abnormal behavior. Hidden risk calculation module: Extracts the deformation vector of the human clothing surface, and based on the phase and amplitude difference between the deformation vector and the skeletal motion, identifies non-natural damping areas and generates hidden foreign object risk values; The hidden risk calculation module generates hidden foreign object risk values ​​by calculating the instantaneous motion vector field of clothing pixels in the human torso area and the translational velocity vector of the corresponding bone nodes. Subtract the translational velocity vector from the instantaneous motion vector field to separate the relative swaying component of the clothing relative to the human torso; Calculate the amplitude of the relative sway component and the phase lag angle of the corresponding component relative to the human gait cycle; If the relative oscillation component amplitude in a local area approaches zero or the phase lag angle shows unnatural locking, it is determined that there is hard object traction or gravity damping in the corresponding area, and the attenuation of the amplitude is normalized and mapped to a hidden foreign object risk value. Postural rigidity calculation module: calculates the symmetry error of the movement trajectory of the left and right limbs, and extracts high-frequency micro-tremor features from the vertical movement sequence of the shoulder and neck nodes to generate the proportion of postural stress risk; The specific process of generating the proportion of postural stress risk in the postural rigidity calculation module is as follows: taking the line connecting the center of the human torso as the axis of symmetry, the spatiotemporal motion trajectory of the left limb skeletal nodes is mirrored and mapped to the right limb coordinate system. The root mean square error of the trajectory overlap of the left and right limbs in a unit gait cycle is calculated, and the root mean square error is normalized to the limb asymmetry coefficient. Vertical displacement data of the shoulder and neck nodes within a continuous time window are extracted, and Fourier transform is used to convert the vertical displacement data from the time domain to the frequency domain. The proportion of spectral energy in the 8Hz to 12Hz physiological tremor frequency band in the total spectral energy is calculated as the high-frequency microtremor coefficient. The limb asymmetry coefficient and the high-frequency microtremor coefficient are weighted and summed to obtain the proportion of postural stress risk that characterizes the target's psychological tension or physiological rigidity. Trajectory wandering degree calculation module: maps the movement trajectory of the identity identifier into a discrete grid sequence, and calculates the discrete entropy value of the trajectory based on the spatiotemporal repeated access frequency; The specific process of calculating the discrete entropy value of the trajectory wandering degree module is to divide the monitoring field of view into a two-dimensional discrete grid of fixed size, and map the coordinate points of the movement trajectory of the identity identifier within a continuous time window into a grid index sequence. By statistically analyzing the frequency of recurrence of each grid index in the sequence and the dwell time of the target in each grid, a spatial distribution probability model of the target's movement trajectory is constructed. The spatial Shannon entropy of the trajectory is calculated based on the spatial distribution probability model. At the same time, the cumulative angle of the direction change of the target motion vector is calculated. The spatial Shannon entropy and the cumulative angle of the direction change are normalized and fused to generate the discrete entropy value of the trajectory, which reflects the degree of disorder of the target's movement route and the degree of wandering in the region. Pathological behavior determination module: Calculates the similarity distance between the current spatiotemporal sequence of the skeleton and the preset pathological gait model, and generates a negative weight exemption coefficient based on the similarity distance; The specific process of generating negative weight exemption coefficients in the pathological behavior judgment module involves pre-collecting skeletal motion data containing various typical pathological gaits and constructing a pathological gait kinematic manifold space that maps to a low-dimensional feature distribution. Calculate the inter-frame Euclidean distance matrix between the real-time spatiotemporal sequence of the current target's skeleton and the standard template in the kinematic manifold space of the pathological gait, and find a time warping path with the minimum cumulative distance in the matrix to perform temporal alignment and eliminate motion rate differences; Calculate the geodesic distance between the projection point of the aligned spatiotemporal sequence of bones in the manifold space and the center of the manifold. Based on a preset exponential decay function, the geodesic distance is mapped to a negative weight exemption coefficient with a negative value. The smaller the geodesic distance, the larger the absolute value of the generated negative weight exemption coefficient.

[0036] The core task of the explicit risk assessment module is to capture and distinguish chaotic and violent motion characteristics. In practice, the module first initiates a full-to-full distance scan of all identified skeleton nodes within the field of view, focusing on monitoring wrist nodes with aggressive attributes and ankle nodes with movement attributes. When the calculated spatial Euclidean distance between any two limb nodes is consistently (e.g., for more than 3 frames) less than a preset contact threshold (e.g., 40 cm), it is determined that these two nodes have entered a potential conflict state, and the spatial range within a preset radius (e.g., 1 meter) with the center of the line connecting these two points as the origin is locked as the interaction area.

[0037] Next, within this interactive area, the real-time velocity vectors of all skeletal nodes are extracted. To accurately quantify the difference between "fighting," "hugging," or "dancing," the system calculates the regional kinetic energy entropy based on statistical mechanics principles. Specifically, it statistically analyzes the variance of the velocity vector amplitude distribution (measuring the variability in motion intensity) and the unit circle dispersion of the direction vector (measuring the disorder in motion direction). When both increase simultaneously, it indicates the presence of disordered, violent motion within the area. The system provides a first-order differential operation on the kinetic energy entropy values ​​over a continuous time series to obtain the rate of evolution of the kinetic energy entropy, reflecting the acceleration trend of the explosive movement.

[0038] in, express The rate of change of kinetic entropy at any given moment reflects the explosive trend of the action; This represents the variance of the magnitude of the velocity vectors of all skeletal nodes within the interaction area; Shannon entropy, representing the distribution of velocity directions, is used to quantify the disorder of motion; This represents the adjustment coefficient for balancing the amplitude and direction weights; This indicates the time interval between video frames.

[0039] If the kinetic entropy evolution rate is detected to show an exponential positive surge in a short period of time (e.g., within 0.5 seconds), and the peak value exceeds the preset biomechanical-based violence limit threshold, it indicates that the explosive force of the action is extremely strong and disordered. In this case, it will be determined that a physical dangerous behavior (such as fighting) has occurred, and a physical danger blocking signal will be generated immediately.

[0040] If an extremely high overall kinetic entropy evolution rate is detected, but no interaction area meeting the contact threshold (i.e., a violent action occurring independently by a single person) is detected between any node pairs within that time window, the system will identify the current behavior as a normal abnormal behavior. This logical distinction ensures that the system can rigidly block violent events while flexibly marking rescue events such as falls.

[0041] Under normal monitoring conditions where no explicit blocking signal is triggered, the latent risk calculation module is run in parallel to analyze the mechanical state under the clothing. First, the instantaneous motion vector field of the clothing surface in the human torso area is calculated pixel by pixel, which is a mixture of the overall translational motion of the clothing with the human body and the flexible swing of the clothing itself. At the same time, the translational velocity vector of the corresponding bone node (usually the center of the hip joint) is read as a rigid reference benchmark.

[0042] By performing vector subtraction, the translational velocity vector of the skeleton is subtracted from every vector in the instantaneous motion vector field. This filters out the overall displacement caused by human movement, separating the relative swaying component of the clothing, which is purely driven by gravity and inertia.

[0043] Subsequently, a time-frequency analysis was performed on the relative oscillation component. In a natural gait cycle, the clothing oscillation should exhibit a pendulum-like sinusoidal wave shape, with a fixed phase lag relative to skeletal movement (determined by the air resistance coefficient of the fabric material). The amplitude of the relative oscillation component and its phase lag angle relative to the gait cycle were calculated in real time.

[0044] If monitoring results show that the relative sway component amplitude of a local area of ​​the torso (such as the right side of the waist) approaches zero (i.e., a dynamic silent zone appears), or its phase lag angle disappears, exhibiting an unnatural locking state completely synchronized with skeletal movement, this indicates that the clothing in that area is supported by the shape or pulled by friction from an internal hard object (such as a gun or knife), losing its original flexible characteristics. Based on the degree of attenuation of the amplitude relative to the natural sway, it is normalized and mapped to a concealed foreign object risk value between 0 and 1.

[0045] The postural rigidity calculation module is responsible for capturing subtle abnormal movements caused by tension, fear, or deliberate camouflage. In practice, the module first constructs a symmetrical human dynamics model. Using the line connecting the center of the human torso (i.e., the line connecting the neck node to the midpoint of the hip) as the geometric axis of symmetry, the coordinates of the motion trajectory of the left limbs (including the left shoulder, left elbow, left wrist, left knee, and left ankle) in a continuous time series are mirrored and mapped to the coordinate system of the right limbs.

[0046] Next, the overlap of the trajectories of the left and right limbs within a unit gait cycle is calculated. A root mean square error (RMSE) algorithm is used to calculate the Euclidean distance between the mirrored left limb node coordinates and the actual right limb node coordinates frame by frame. A larger error value indicates greater incoordination of bilateral limb movements (e.g., stiffness caused by a foreign object lodged in one side). The RMSE is then normalized to a limb asymmetry coefficient between 0 and 1 using a nonlinear function (such as the sigmoid function).

[0047] in, The root mean square error of the limb movement trajectory is the original asymmetric metric. This indicates the total number of limb skeletal nodes involved in the calculation (e.g., 5 nodes in total: shoulder, elbow, wrist, knee, and ankle). Indicates the leftmost Real-time coordinate vectors of each skeletal node; Indicates the right-hand side Real-time coordinate vectors of each skeletal node; A function representing a mirror image of the human torso about its central axis; This represents the Euclidean distance operation.

[0048] Simultaneously, high-frequency micro-motion analysis was performed on the shoulder and neck nodes. To capture muscle tremors that are imperceptible to the naked eye, the system recorded the vertical displacement data of these nodes at frequencies that satisfy the Nyquist sampling theorem. The displacement waveform in the time domain was converted into a power spectral density map in the frequency domain using a Fast Fourier Transform. Special attention was paid to the specific frequency band of 8Hz to 12Hz, which is the typical physiological micro-tremor frequency produced by muscles under high stress or tension. The proportion of spectral energy within this frequency band in the total spectral energy was calculated to generate high-frequency micro-tremor coefficients.

[0049] Finally, the weighted sum of the limb asymmetry coefficient and the high-frequency micro-tremor coefficient is used to output the proportion of postural stress risk. This indicator can reflect the abnormal state of the target who appears to be walking calmly but is actually stiff and tense.

[0050] The trajectory loitering degree calculation module aims to identify regional loitering behavior with unclear intentions. First, the ground plane of the monitored field of view is divided into a two-dimensional discrete grid of fixed size (e.g., 1 meter × 1 meter). For each global identity, the sequence of its motion trajectory coordinates within a continuous time window (e.g., 30 seconds) is mapped to the corresponding grid index sequence.

[0051] By statistically analyzing the recurrence frequency of each grid index in the sequence and summing the target's dwell time within each grid, a spatial distribution probability model of the target's trajectory is constructed. Based on the Shannon entropy formula in information theory, the spatial Shannon entropy of this probability model is calculated. This entropy value reflects the determinism of the target's spatial distribution: targets traversing in a straight line have a uniform entropy distribution, while targets staying at a fixed point have extremely low entropy values.

[0052] in, The spatial Shannon entropy of a trajectory is used to quantify the degree of dispersion of a spatial distribution; This represents the total number of discrete grids divided into the monitored field of view; Indicates the first One grid index; Indicates the target is in the first place. The probability ratio of dwell time within a grid to the total observation time satisfies .

[0053] To accurately distinguish between normal loitering (such as waiting) and suspicious lingering, it is necessary to calculate the angle between adjacent frames of the target's motion vector and obtain the cumulative angle of directional change within the time window. Lingering behavior is usually manifested as frequent back-and-forth, circling, and turning within a limited area, which will lead to a significant increase in the cumulative angle value.

[0054] The spatial Shannon entropy is normalized and fused with the cumulative angle of directional change. High-frequency back-and-forth motion (high cumulative angle) combined with specific regional limitation distribution (i.e., non-linear crossing) will generate extremely high trajectory discrete entropy values, thereby effectively marking suspicious targets wandering aimlessly in the region.

[0055] To avoid misclassifying individuals with disabilities or neurological disorders as abnormal targets exhibiting rigid postures, the pathological behavior detection module introduces a deep filtering mechanism based on manifold learning. First, the module pre-configures a pathological gait database, which collects and labels skeletal motion data from various typical pathologies, including Parkinson's disease (characterized by a hurried, shuffling gait), stroke hemiplegia (characterized by a dragging, circling gait), and sequelae of poliomyelitis. Given the non-linear characteristics of high-dimensional skeletal motion data, the module constructs a low-dimensional kinematic manifold space for pathological gait. In this space, similar pathological patterns cluster on specific manifold surfaces, forming topologically structured distribution clusters.

[0056] In real-time monitoring, the module acquires the real-time spatiotemporal sequence of the current target's skeleton. Considering that patients' movement speed is typically significantly slower than normal or has an unstable rhythm, direct linear matching would lead to substantial computational errors. Therefore, the module implements a dynamic temporal warping strategy. It calculates the inter-frame Euclidean distance matrix between the real-time skeleton sequence and the standard pathological template in manifold space, and then searches for a temporally warped path with the minimum cumulative cost within this matrix. By non-linearly stretching or compressing the sequence along this path, elastic temporal alignment is achieved, effectively eliminating interference caused by differences in movement speed.

[0057] After alignment, the module projects the target's skeletal sequence onto a low-dimensional manifold space and calculates the geodesic distance between the projection point and the center of the pathological manifold cluster. Unlike traditional Euclidean distance, the geodesic distance is calculated along the manifold surface, which can more realistically reflect the similarity between the target gait and the standard pathological gait in terms of their intrinsic dynamic structure.

[0058] Based on a preset exponential decay function, geodesic distances are mapped to negative weighted exemption coefficients with negative values. When the geodesic distance is extremely small (i.e., highly consistent with a certain pathological gait), a negative coefficient with a large absolute value is generated (e.g., -0.8); as the distance increases, the coefficient rapidly decays to zero.

[0059] in, This represents the negative weight exemption coefficient generated; the value is negative. This represents the geodesic distance between the current skeletal sequence projection point and the pathological manifold center; The maximum exemption strength constant (e.g., 0.8) determines the maximum absolute value of the negative weight when the distance is 0; This represents the decay rate parameter, which controls how quickly the exemption coefficient decays as similarity decreases (distance increases).

[0060] Potential risk assessment module: integrates the risk value of concealed foreign objects, the proportion of postural stress risk, and the trajectory discrete entropy value to generate the initial hidden risk coefficient corresponding to the identity identifier; The specific process of generating the initial latent risk coefficient in the potential risk assessment module involves performing interval mapping normalization on the value of the hidden foreign object risk, the proportion of body stress risk, and the trajectory discrete entropy value, respectively. The normalized values ​​are used as independent orthogonal components to construct a three-dimensional feature vector representing the latent state of the target. The Euclidean modulus of the three-dimensional feature vector is calculated and used as the initial latent risk coefficient to quantify the potential threat level of the corresponding identity identifier. Overall Risk Assessment Module: Upon receiving a physical hazard blocking signal, it directly locks the alarm status; it comprehensively assesses the latent hazard coefficient and the object's hazard level after correction by the negative weight exemption coefficient, and marks the corresponding identity as an abnormal risk target and outputs a risk warning signal when the total risk value continues to be higher than the threshold. The overall risk assessment module performs a comprehensive assessment and early warning process, establishes a listening channel for physical hazard blocking signals, and locks the risk status of the current identity to the highest level and triggers an alarm upon detecting a physical hazard blocking signal. When no physical danger blocking signal is detected, the object detection result associated with the current identity is obtained, and the pre-set dangerous item level table is queried to determine the corresponding object danger weight; The initial implicit risk coefficient and the negative weight exemption coefficient are algebraically superimposed and corrected, and the corrected result is added to the object risk weight to obtain the real-time total risk value. A fixed-length sliding time window queue is constructed to store the real-time total risk value of historical frames. If the real-time total risk value of all data frames in the corresponding queue is higher than the preset composite warning threshold, an anomaly is confirmed, the identity is marked as an abnormal risk target, and a warning signal containing the anomaly type and evidence fragments is output.

[0061] After each submodule completes feature extraction, the potential risk assessment module receives three core indicators from upstream: the value of the hidden foreign object risk, the proportion of postural stress risk, and the trajectory discrete entropy value. Since these indicators have different physical dimensions and numerical ranges (for example, the risk value is 0-1, while the entropy value may be greater than 1), they need to be normalized by interval mapping to uniformly map them to the standard dimensionless interval [0,1].

[0062] Next, a three-dimensional latent risk feature space is established, and the three normalized values ​​are used as independent orthogonal components on the X, Y, and Z axes, respectively, as a three-dimensional feature vector describing the current latent state of the target.

[0063] Calculate the Euclidean modulus (L2 norm) of this three-dimensional feature vector. Geometrically, this modulus represents the distance of the target from the "absolute safety origin" in the risk space. Use this modulus as the initial latent risk coefficient, which comprehensively reflects the cumulative threat level of the target across three dimensions: carrying foreign objects, psychological stress, and unusual behavior.

[0064] in, This represents the initial latent risk factor; Represents a three-dimensional latent risk feature vector; This represents the normalized value for the risk of hidden foreign objects. This represents the normalized percentage of postural stress risk. This represents the normalized discrete entropy value of the trajectory.

[0065] The overall risk assessment module establishes a high-priority physical hazard blocking signal monitoring channel. Once a blocking signal from the explicit risk judgment module is detected (such as a fight), the system immediately terminates subsequent calculations through an interrupt mechanism, directly locks the risk status of the current identity to the highest level, and triggers an immediate alarm report to the monitoring personnel.

[0066] The system will continue to make comprehensive judgments when no explicit blocking signal is received. First, it will retrieve the object detection results associated with the current identity and query the preset dangerous item level table (e.g., firearm weight = 0.8, controlled knife weight = 0.5, ordinary item weight = 0) to determine the object danger weight of the current target.

[0067] The initial implicit risk coefficient is algebraically added to the negative weight exemption coefficient. This achieves logical risk hedging; even if the target's implicit risk is high (e.g., gait stiffness due to disability), if its pathological exemption coefficient is also high (confirmed as pathological gait), the risk value will be significantly reduced or even reduced to zero after adding the two. Then, the corrected result is added to the object's risk weight to obtain the real-time total risk value.

[0068] in, This represents the total real-time risk value ultimately used for sliding window validation; This represents the negative weighted exemption coefficient (this value is negative) generated by the pathological behavior determination module. This represents the object hazard weight obtained by looking up the table based on the object detection results; This represents the modified implicit risk lower bound constraint, preventing the risk value from becoming negative due to excessive exemptions (although mathematically permissible, the risk is usually non-negative in a physical sense).

[0069] To eliminate false alarms in a single frame caused by video noise or instantaneous movements, the module constructs a sliding time window queue of fixed length (e.g., the number of frames corresponding to 5 consecutive seconds). The calculated real-time total risk value for each frame is pushed into the queue. Only when the real-time total risk value of all data frames in the queue is consistently higher than a preset composite warning threshold (e.g., 0.75) is an anomaly finally confirmed. At this point, the system marks the identified person as an anomalous risk target, automatically extracts video clips before and after the anomaly as evidence, and outputs a warning signal containing the anomaly type (e.g., "armed and nervous") to the monitoring terminal.

[0070] In summary, this invention enables rapid alerts after violent incidents by monitoring the kinetic entropy evolution rate of the interaction areas of limb nodes. It also overcomes the technical blind spot of traditional vision in identifying concealed foreign objects by utilizing the phase difference and amplitude attenuation analysis of clothing optical flow and skeletal movement. Furthermore, by constructing a mirror-symmetric model of the limbs and extracting the 8-12Hz physiological micro-tremor features of the shoulder and neck, it keenly captures microscopic postural abnormalities caused by psychological stress. Combined with negative weighted exemption coefficients generated based on manifold learning and dynamic time warping techniques, it eliminates gait interference caused by physiological diseases. Simultaneously, it integrates appearance and gait features to maintain identity consistency across all scenarios and uses sliding time windows and multi-factor algebraic superposition logic to filter out video noise. The synergistic effect of these technologies constructs a real-time abnormal behavior identification and intelligent early warning system that possesses both the ability to effectively block overt violent behavior and the ability to penetrate disguises to identify concealed suspicious individuals, while also considering humanitarian concerns and exhibiting extremely high robustness.

[0071] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis, characterized in that, It consists of a set of highly collaborative functional modules, specifically including: Skeleton attachment module: Locates the human body region in the image frame of the synchronized video data stream, extracts the feature response extreme points within the region, connects the feature response extreme points into a skeleton topology according to human kinematic constraints, and converts the pixel coordinates into normalized coordinates; Cross-view labeling module: It integrates human appearance texture and skeletal motion cycle to generate feature vectors, and based on the matching results of the feature vectors and the global trajectory library, it unifies the identity of the same target under different viewpoints; Explicit risk assessment module: calculates the kinetic energy entropy evolution rate of the interaction region of the skeleton limb nodes, and generates a physical danger blocking signal when the kinetic energy entropy evolution rate exceeds the preset safety boundary; Potential risk assessment module: integrates the risk value of concealed foreign objects, the proportion of postural stress risk, and the trajectory discrete entropy value to generate the initial hidden risk coefficient corresponding to the identity identifier; Overall risk assessment module: Upon receiving a physical hazard blocking signal, it directly locks the alarm status; it comprehensively assesses the latent hazard coefficient and the object's hazard level after correction by the negative weight exemption coefficient, and when the total risk value continues to exceed the threshold, it marks the corresponding identity as an abnormal risk target and outputs a risk warning signal.

2. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 1, characterized in that, The system also includes the following modules: Data alignment module: It uses a built-in clock to uniformly calibrate multiple monitoring devices and align the timestamps of multiple video streams to form a synchronized video data stream; Hidden risk calculation module: Extracts the deformation vector of the human clothing surface, and based on the phase and amplitude difference between the deformation vector and the skeletal movement, identifies non-natural damping areas and generates hidden foreign object risk values; Postural rigidity calculation module: calculates the symmetry error of the movement trajectory of the left and right limbs, and extracts high-frequency micro-tremor features from the vertical movement sequence of the shoulder and neck nodes to generate the proportion of postural stress risk; Trajectory wandering degree calculation module: maps the movement trajectory of the identity identifier into a discrete grid sequence, and calculates the discrete entropy value of the trajectory based on the spatiotemporal repeated access frequency; Pathological behavior determination module: Calculates the similarity distance between the current spatiotemporal sequence of the skeleton and the preset pathological gait model, and generates a negative weight exemption coefficient based on the similarity distance.

3. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 1, characterized in that, The specific process of locating and extracting the skeleton topology by the skeleton attachment module includes: Perform multi-scale feature extraction and convolution operations on the image frame to identify regions with human appearance texture features and generate human body bounding boxes; Calculate the feature response intensity of pixel data within the human body bounding box to key anatomical parts of the human body, generate a probability response heatmap including the top of the head, neck, shoulder, elbow, wrist, hip, knee and ankle, and select the local response extreme points in the heatmap as key skeletal points. A local Cartesian coordinate system is constructed with the midpoint of the line connecting the left and right hip joints as the origin and the pixel height of the human body bounding box as the unit length. The absolute pixel coordinates of the key bone points are projected into the local Cartesian coordinate system to generate the skeleton topology that eliminates the difference between shooting distance and viewpoint scale.

4. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 1, characterized in that, The specific process of unifying identity identification through the cross-view domain tagging module includes: Extract the color distribution histogram and texture gradient features of the human body region, and calculate the gait interaction frequency and stride features of the bipedal nodes in the skeleton topology. Concatenate and stitch the color distribution histogram, texture gradient features and gait features into a multidimensional identity description vector. Calculate the cosine similarity between the current multidimensional identity description vector and the centroids of each stored feature vector set in the global trajectory database; If the calculated highest cosine similarity exceeds the preset identity association threshold, the current multidimensional identity description vector is assigned to the corresponding feature vector set, and the centroid of the feature vector set is updated as the identity matching template for the corresponding target in the next moment. If the calculated highest cosine similarity is lower than the identity association threshold, the current target is determined to be a newly emerging target and a new globally unique identity is assigned. A new feature vector set is initialized in the global trajectory database.

5. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 1, characterized in that, The specific process by which the explicit risk assessment module generates a physical danger blocking signal includes: Real-time monitoring of wrist and ankle nodes of the human skeleton within the field of view; when the spatial Euclidean distance between skeleton nodes is less than a preset contact threshold, the corresponding spatial range is locked as the interactive area. Calculate the velocity vectors of all nodes within the interaction region, calculate the regional kinetic energy entropy based on the amplitude distribution variance and direction dispersion of the velocity vectors, and obtain the kinetic energy entropy evolution rate by differentiating the regional kinetic energy entropy over a continuous time series. The kinetic energy entropy evolution rate is compared with a preset biomechanical violence limit threshold. If the kinetic energy entropy evolution rate shows a positive surge and the peak value exceeds the violence limit threshold, it is determined that a physical dangerous behavior has occurred, and the physical danger blocking signal is generated. Furthermore, if the kinetic entropy evolution rate is detected to exceed the violent limit threshold, but no interactive area satisfying the contact threshold is formed between the skeleton nodes, the current behavior is identified as a normal abnormal behavior.

6. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 2, characterized in that, The specific process by which the latent risk calculation module generates the hidden foreign object risk value includes: Calculate the instantaneous motion vector field of clothing pixels within the human torso region, and calculate the translational velocity vector of the corresponding bone nodes; Subtract the translational velocity vector from the instantaneous motion vector field to separate the relative swaying component of the clothing relative to the human torso; Calculate the amplitude of the relative sway component and the phase lag angle of the corresponding component relative to the human gait cycle; If the relative oscillation component amplitude in a local area is detected to be close to zero or the phase lag angle is found to be unnaturally locked, it is determined that there is hard object traction or gravity damping in the corresponding area, and the attenuation degree of the amplitude is normalized and mapped to a hidden foreign object risk value. Furthermore, the specific process by which the postural rigidity calculation module generates the proportion of postural stress risk includes: Using the line connecting the center of the human torso as the axis of symmetry, the spatiotemporal motion trajectory of the left limb skeletal nodes is mirrored and mapped to the right limb coordinate system. The root mean square error of the trajectory overlap between the left and right limbs within a unit gait cycle is calculated, and the root mean square error is normalized to the limb asymmetry coefficient. Vertical displacement data of the shoulder and neck nodes within a continuous time window are extracted, and the vertical displacement data are transformed from the time domain to the frequency domain using Fourier transform. The proportion of spectral energy in the 8Hz to 12Hz physiological tremor frequency band in the total spectral energy is calculated as the high-frequency microtremor coefficient. The limb asymmetry coefficient and the high-frequency microtremor coefficient are weighted and summed to obtain the proportion of postural stress risk that characterizes the target's psychological tension or physiological rigidity.

7. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 2, characterized in that, The specific process by which the trajectory wandering degree calculation module calculates the discrete entropy value of the trajectory includes: The monitoring field of view is divided into a two-dimensional discrete grid of fixed size, and the coordinate points of the motion trajectory of the identity identifier within a continuous time window are mapped to a grid index sequence. By statistically analyzing the frequency of recurrence of each grid index in the sequence and the dwell time of the target in each grid, a spatial distribution probability model of the target's movement trajectory is constructed. The spatial Shannon entropy of the trajectory is calculated based on the spatial distribution probability model. At the same time, the cumulative angle of the direction change of the target motion vector is calculated. The spatial Shannon entropy and the cumulative angle of direction change are normalized and fused to generate a discrete entropy value of the trajectory that reflects the degree of disorder of the target's movement route and the degree of wandering in the region.

8. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 1, characterized in that, The specific process by which the potential risk assessment module generates the initial latent hazard coefficient includes: The hidden foreign object risk value, the proportion of body stress risk, and the trajectory discrete entropy value are respectively subjected to interval mapping normalization processing. The normalized values ​​are used as independent orthogonal components to construct a three-dimensional feature vector representing the hidden state of the target. The Euclidean modulus of the three-dimensional feature vector is calculated and used as the initial hidden risk coefficient to quantify the potential threat level of the corresponding identity identifier.

9. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 2, characterized in that, The specific process by which the pathological behavior determination module generates the negative weight exemption coefficient includes: Skeletal motion data containing various typical pathological gaits were pre-collected, and a pathological gait kinematic manifold space mapping a low-dimensional feature distribution was constructed. Calculate the inter-frame Euclidean distance matrix between the real-time spatiotemporal sequence of the current target's skeleton and the standard template in the pathological gait kinematic manifold space, and find a time warping path with the minimum cumulative distance in the matrix to perform temporal alignment and eliminate motion rate differences; Calculate the geodesic distance between the projection point of the aligned spatiotemporal sequence of bones in the manifold space and the center of the manifold. The geodesic distance is mapped to a negative weight exemption coefficient based on a preset exponential decay function, wherein the smaller the geodesic distance, the larger the absolute value of the generated negative weight exemption coefficient.

10. The real-time identification and intelligent early warning system for abnormal behavior based on structured video stream analysis according to claim 1, characterized in that, The specific process by which the overall risk assessment module performs comprehensive assessment and early warning includes: Establish a monitoring channel for the physical hazard blocking signal. Upon detecting the physical hazard blocking signal, lock the risk status of the current identity to the highest level and trigger an alarm. If no physical hazard blocking signal is detected, obtain the object detection result associated with the current identity identifier, and query the preset dangerous item level table to determine the corresponding object hazard weight; The initial latent hazard coefficient and the negative weight exemption coefficient are algebraically superimposed and corrected, and the corrected result is added to the object hazard weight to obtain the real-time total risk value; A fixed-length sliding time window queue is constructed to store the real-time total risk value of historical frames. If the real-time total risk value of all data frames in the corresponding queue is higher than the preset composite warning threshold, an anomaly is confirmed, the identity identifier is marked as an abnormal risk target, and a warning signal containing the anomaly type and evidence fragments is output.

Images