Computer vision-based park intelligent monitoring system

Abstract

The present invention discloses an intelligent monitoring system for a park based on computer vision, which relates to the technical fields of computer vision and intelligent monitoring. It collects video image frames of multiple cameras and constructs a video stream atlas with cross-camera spatial synchronization; extracts static structure features and dynamic change features through semantic segmentation and optical flow estimation to generate a multi-modal scene structure feature vector; constructs an inter-frame micro-action evolution path atlas, extracts the micro-change action set of the target and the candidate sequence of abnormal behaviors; calculates the co-occurrence frequency of micro-actions and constructs a multi-target behavior coupling atlas; classifies and identifies abnormal behaviors based on a behavior scoring model and triggers a linkage response; the present invention realizes the accurate capture and risk identification of fine-grained behavior changes, and improves the intelligence, real-time performance and safety prevention and control capabilities of the monitoring system in a complex park environment.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent monitoring technology, specifically to a computer vision-based intelligent monitoring system for industrial parks. Background Technology

[0002] As industrial parks continue to expand, traditional video surveillance systems face numerous challenges in terms of the real-time nature, accuracy, and proactivity of event detection. Current park monitoring primarily relies on a combination of numerous cameras and manual patrols, using target detection and behavior recognition algorithms to identify abnormal behavior. However, this approach has limited ability to capture "micro-movements," "abnormal interactive behaviors," and "precursor states" in emergencies, especially in multi-person scenarios, heavily obstructed environments, or low-light conditions. These systems are prone to false positives or false negatives, leading to delayed emergency response and failing to meet the actual needs of high-security parks.

[0003] In addition, existing systems generally use single-frame image analysis technology, which lacks the ability to model the temporal correlation between multiple frames, making it difficult to accurately predict the evolution trend of target behavior. In terms of target identification, traditional methods rely on facial or body features, which are easily affected by changes in angle and interference from occlusions, and have significant limitations.

[0004] Even more serious is that in high-risk areas such as hazardous chemical storage areas, high-voltage substations, and semi-enclosed underground parking lots, some early warning signs of safety accidents are only reflected in very brief changes in characteristics in individual frames, such as abnormal lingering of personnel, an increase in the frequency of abnormal turning around, or abnormal placement of items. Traditional systems have difficulty in effectively modeling and judging such "weak suspicious behaviors," and thus it is difficult to achieve early warning and source tracing of potential risks.

[0005] Therefore, there is an urgent need for an intelligent monitoring method that can integrate computer vision, multimodal temporal modeling, and micro-motion aggregation analysis to improve the system's ability to identify fine-grained abnormal behaviors and solve the problems of perception blind spots and early warning delays in existing systems in high-security monitoring scenarios in parks. Summary of the Invention

[0006] The purpose of this invention is to provide a computer vision-based intelligent monitoring system for industrial parks to address the shortcomings of the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a computer vision-based intelligent monitoring system for industrial parks, comprising: Video acquisition module: Collects video image frame sequences from multiple surveillance cameras within the park at different time periods, and constructs a video stream map that is spatially synchronized across cameras; Feature extraction module: Based on the video stream graph, a region segmentation algorithm is used to extract static structural features and dynamic change features from each image frame to generate a multimodal scene structural feature vector; Micro-motion modeling module: Input the multimodal scene structure feature vector into the temporal graph convolutional network to construct an inter-frame micro-motion evolution path map and extract the target's micro-change action set and abnormal behavior candidate sequence in consecutive frames; Frequency map construction module: Based on the inter-frame micro-motion evolution path map, calculate the co-occurrence frequency map of micro-motions of each target within the historical time window, and construct a multi-target behavior coupling map; Abnormal Behavior Analysis Module: Based on the multi-target behavior coupling map, construct an abnormal behavior scoring model, perform risk scoring and clustering classification on abnormal interaction behaviors between targets, and identify potential security risk behaviors; Intelligent linkage acquisition module: Based on the risk score results, it automatically generates a microscopic observation feature coordinate system in high-risk areas, and controls the camera to adjust its angle and focal length, performing high-frequency image acquisition and local fine-grained recognition analysis on key targets; Response module: Compares the analysis results with the preset security behavior template library, outputs the abnormal behavior level assessment results, and triggers the response strategy.

[0008] Preferably, constructing a cross-camera spatially synchronized video stream atlas includes the following steps:

[0009] The system collects a sequence of video image frames continuously captured by fixed-view surveillance cameras deployed in different areas of the park within a preset time period; timestamps are aligned between different video image frames based on the spatial location information and time synchronization signals of each camera; target correspondence is calculated for image frames within the coverage area of ​​adjacent cameras using a feature point matching algorithm, and a target trajectory association model is constructed between cameras; a video stream map under unified spatial coordinates is generated based on the continuity of the target trajectory and scene structure information.

[0010] Preferably, the method of extracting static structural features and dynamic change features from each image frame using a region segmentation algorithm includes the following steps: The image frames in the video stream atlas are processed for region segmentation. A pre-trained semantic segmentation model is used to classify road regions, building structure regions and activity regions at the pixel level to generate a static structural region distribution map. Based on the static structural region distribution map, inter-frame difference calculation and optical flow field estimation are performed on continuous image frames to extract pixel motion change information of the target in the active region and generate dynamic change feature map. Based on the dynamic change feature map and the static structure region distribution map, the target region is subjected to feature weighted fusion calculation to form a scene structure feature set of unified dimension; The scene structure feature set is vectorized and encoded in chronological order to generate a multimodal scene structure feature vector.

[0011] Preferably, constructing the inter-frame micro-motion evolution path map includes the following steps: The multimodal scene structure feature vectors are used to construct a node graph structure in chronological order, where each frame of the image corresponds to a graph node, and initial edge connections are established between nodes based on their chronological order. Based on the feature similarity and time interval between nodes, a Gaussian kernel function is used to calculate the edge weights and construct a weighted time sequence graph structure. The weighted temporal graph is input into a joint temporal graph convolutional network containing graph convolutional layers and gated recurrent units to extract the micro-action state transition features between nodes; Based on the difference in feature change rate and action vector direction between nodes, an inter-frame micro-motion evolution path map is generated.

[0012] Preferably, the extraction of the target's micro-change action set and anomalous behavior candidate sequence in consecutive frames includes the following steps: The difference is calculated on the target feature vectors corresponding to adjacent graph nodes in the inter-frame micro-motion evolution path map. The changes in the target's spatial position, velocity direction and regional semantics at each moment are extracted to form an inter-frame motion difference vector. Based on the inter-frame motion difference vector, a sliding time window clustering algorithm is used to identify micro-motion segments with similar change patterns in a continuous frame sequence, and to construct a set of micro-change motions of the target. For each micro-motion segment, calculate its rate of change curve and trajectory direction offset value, and compare the deviation with the historical normal behavior template; if the deviation exceeds the set threshold, mark the corresponding frame sequence as an abnormal behavior candidate sequence.

[0013] Preferably, constructing a multi-objective behavior coupling map includes the following steps: Based on the inter-frame micro-motion evolution path map, the micro-change action set of each target in the continuous time series is extracted, and the time overlap interval of different target micro-motion segments is counted within a preset historical time window to calculate the number of micro-motion co-occurrences between targets. Based on the co-occurrence frequency and the change in spatial distance between targets, normalization is used to calculate the co-occurrence frequency weight of micro-actions between targets, and a correlation weight matrix between targets is constructed. Based on the association weight matrix, targets with co-occurrence frequencies higher than a set threshold are connected to generate a multi-target behavior association graph structure. Based on the connection strength and temporal continuity between each objective, a multi-objective behavior coupling map is formed.

[0014] Preferably, the risk scoring and clustering classification of abnormal interactions between targets includes the following steps: Extract the temporal continuity, co-occurrence frequency weights, and spatial proximity features of each connection edge in the multi-objective behavior coupling graph, and construct a multi-dimensional behavior interaction feature vector. The feature vector is input into the behavior anomaly scoring model trained by supervised learning, and the risk score value of each group of target interaction behaviors is output. Based on the risk score, density clustering algorithm is used to cluster and classify behavioral samples, dividing them into low-risk, suspicious and high-risk behavior categories; Target pairs in high-risk categories are identified and, combined with their spatiotemporal trajectories and semantic regions, are marked as potential security risk behaviors.

[0015] Preferably, the linkage control of the camera to adjust its angle and focus includes the following steps:

[0016] Based on the location of high-risk targets and their semantic regions output by the abnormal behavior scoring model, a microscopic observation feature coordinate system is constructed, which includes information on target location, movement trend, and surrounding structure. Based on the target's velocity direction vector and regional occlusion conditions in the coordinate system, a camera group with the best field of view coverage is selected. The selected cameras are controlled to adjust their pitch angle, horizontal rotation angle, and focal length parameters in real time so that their main field of view is aligned with the target's movement path.

[0017] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0018] 1. This invention effectively improves the ability to identify fine-grained abnormal behaviors in complex park environments by introducing multimodal scene structural feature modeling, inter-frame micro-motion evolution path map construction, and multi-target behavior coupling analysis. Especially in multi-person scenes, occluded environments, or low-dynamic behavior backgrounds, it can accurately extract micro-motion differences, behavior correlation frequency, and temporal continuity features, significantly reducing false alarm and false negative rates, and achieving early warning and evolution trend perception of abnormal behaviors.

[0019] 2. This invention constructs a microscopic observation feature coordinate system based on risk scoring, links camera angle and focal length parameters for automated image acquisition, and combines a behavior template comparison mechanism for hierarchical anomaly assessment and response strategy triggering. This not only improves the system's efficiency in dynamically handling high-risk targets, but also realizes the intelligent transformation of park monitoring from "passive observation" to "active intervention," possessing the advantages of high precision, high timeliness, and scalability. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] For examples, please refer to Figure 1 As shown, the computer vision-based intelligent monitoring system for industrial parks described in this embodiment includes: Video acquisition module: Acquires video image frame sequences from multiple surveillance cameras within the park at different time periods, constructing a cross-camera spatially synchronized video stream map. Specifically, it includes the following steps: Fixed-angle surveillance cameras deployed at park entrances, corridors, perimeter walls, and key material storage locations collect video data at a rate of 25 frames per second, with each frame measuring 1920×1080 pixels. The cameras connect to a unified control server via a local time synchronization protocol, embedding millisecond-accurate timestamps into each frame. The data acquisition process is continuous for a set time period, such as 2 hours, to cover the entire human activity cycle. All image frames are named with "camera number + timestamp" and cached chronologically on edge computing nodes. The acquired data is in H.264 encoded video stream format, subsequently decoded frame-by-frame into raw RGB image frames for later processing.

[0024] Since there are slight time discrepancies between image frames captured by different cameras, a global time calibration method is used to uniformly synchronize all cameras to ensure the accuracy of subsequent frame-level comparison. The specific method is as follows: Using Network Time Protocol (NTP) or GPS clock modules, the time error of all camera systems is controlled within ±10 milliseconds; Set the synchronization threshold ΔT to 100 milliseconds, which means aligning all image frames captured by the cameras to the nearest 100-millisecond window according to their timestamps; If a camera is missing a frame in the current time window, a fitted frame is generated through inter-frame interpolation (based on linear interpolation between consecutive frames) to maintain temporal continuity. The final output is a time-synchronized video frame sequence group for use in cross-camera image matching.

[0025] For time-synchronized image frames, target correspondence calculations are performed between images to establish spatial matching relationships for the same target from different camera perspectives. The implementation steps are as follows: The Scale Invariant Feature Transform (SIFT) algorithm was used to extract key feature points in each frame of the image, with an average of more than 500 points extracted per frame; Bidirectional feature matching is performed on image pairs with overlapping visual areas. The matching distance is measured using Euclidean distance, and the matching threshold is set to 0.6 times the nearest neighbor ratio (Lowe's ratio test). The RANSAC algorithm is used to calculate the homography matrix for candidate matching point pairs to eliminate false matches; If the number of valid matches is greater than 50 and the fitting error is less than 3 pixels, then the image pair is determined to have a target correspondence. Cluster analysis of the spatial location of the same target in multiple frames (DBSCAN algorithm, distance threshold of 30 pixels, minimum number of samples of 5). The spatial correspondence of the target under each camera is identified using the clustering results and recorded in the form of "target number + camera number + image frame number + coordinate point"; Generate a list of target trajectories across cameras to describe the continuous movement path of the target from different viewpoints.

[0026] To achieve spatial uniformity of data from all cameras, the image coordinate system in each camera is converted into a unified world coordinate system, and the following calculation process is performed: Each camera is deployed by measuring its installation height, pitch angle, azimuth angle and offset distance from the origin using a total station or lidar, and establishing an external parameter matrix, including the rotation matrix R and translation vector T. For any pixel p in the image, first convert it to three-dimensional coordinates in the camera coordinate system using the intrinsic parameter matrix, and then perform three-dimensional spatial transformation using the extrinsic parameter matrix to map it to a point P in the unified world coordinate system. Using the continuous position points of the target in a unified world coordinate system as nodes and the timestamp as an index, a three-dimensional space + time trajectory graph structure is constructed; Each trajectory edge contains attributes such as target ID, start and end cameras, start and end times, displacement vector, and velocity change; Finally, a video stream atlas with "target-time-space" as the three-dimensional main axis is constructed as the basic input structure for downstream micro-motion modeling and anomaly recognition.

[0027] This implementation plan ensures that the spatial trajectory of the same target is uniformly represented from different perspectives, providing accurate data support for the system to realize behavioral pattern analysis and intelligent response decision-making in non-continuous images across cameras.

[0028] Feature extraction module: Based on the video stream graph, a region segmentation algorithm is used to extract static structural features and dynamic change features from each image frame, generating a multimodal scene structural feature vector.

[0029] To achieve joint modeling of static and dynamic features in park video images, and to extract multimodal scene structural feature vectors with spatial structure and behavioral representation capabilities, the following steps are performed:

[0030] Each frame of the constructed video stream atlas is sequentially input into a semantic segmentation neural network pre-trained on a city scene dataset (such as Cityscapes or ADE20K). The specific model used is the DeepLabV3+ network structure, with ResNet-101 as the backbone encoder. Different dilation rates (such as 6, 12, and 18) are set in each dilated convolutional layer to obtain multi-scale receptive fields.

[0031] After performing convolutional feature extraction, dilated spatial pyramid pooling, and upsampling reconstruction on the input image, the semantic category probability value corresponding to each pixel is output, and the category corresponding to the maximum value is the final classification label of the pixel.

[0032] This embodiment unifies semantic categories into the following categories: Building structure area (label number 1). Road area (label number 2); Green areas (label number 3); Activity area (label number 4); Other unidentifiable areas (label number 0).

[0033] The output is a two-dimensional matrix with the same dimensions as the original image, called the static structure region distribution map, where each pixel value is the category number mentioned above. This distribution map is used to restrict dynamic analysis to occur only in active regions, thereby improving computational efficiency and data accuracy.

[0034] For any two consecutive frames in the video stream graph (e.g., frame t and frame t+1), perform the following two processes in their active region to extract dynamic change features:

[0035] Convert the image to grayscale; calculate the difference between the grayscale values ​​of the pixel at position (x, y) in the two frames, where the difference is defined as... If D(x,y) is greater than the set threshold of 30, the pixel position is considered to have changed significantly and is recorded as a dynamic region pixel. In the above difference map, pixels outside the active region are forcibly set to 0 values ​​to avoid background noise interference.

[0036] Within the dynamic region marked by the aforementioned difference map, sparse optical flow calculation based on the pyramid Lucas-Kanade algorithm is performed. This algorithm selects corner points (e.g., Harris corner points) with feature intensities higher than a threshold (e.g., 30) and stable edge responses on each layer of the image pyramid, and tracks the positional changes of these corner points in the next frame through image window matching.

[0037] Each successfully tracked point outputs a two-dimensional displacement vector (Δx, Δy), representing the direction and magnitude of the pixel's movement. All vectors are combined to form an optical flow field map, called a dynamic change feature map, which is used to express the directionality and continuity of the target's micro-movements.

[0038] To integrate static structural information with dynamic behavioral information, the following operations are performed on the targets in each active region: Locating the target region: Perform connected component analysis on regions with a movement amplitude greater than 0.5 pixels in the optical flow field map, and extract bounding boxes as candidate target regions; Extract static structural features: Within the target bounding box, count the pixel percentage of each semantic category in the static structural region distribution map to form a set of structural feature values ​​(e.g., activity area accounts for 70%, road area accounts for 20%, and green area accounts for 10%). Extracting dynamic feature information: Statistically analyze the average magnitude and direction distribution histogram of all optical flow vectors within the target area to obtain motion intensity and directional features.

[0039] Weighted fusion: Static structural features and dynamic features are fused through linear weighting. The fusion formula is: Feature vector F = α × Fstatic + β × Fdynamic, where α = 0.4 and β = 0.6, which is used to highlight the dominant role of dynamic features in abnormal behavior recognition.

[0040] Each target ultimately generates a scene structure feature sub-vector in a unified format. This vector has a length of 128 dimensions and includes fields such as static semantic proportion, average movement amplitude, and directional statistics.

[0041] All target feature sub-vectors extracted from each frame are arranged chronologically to construct a time-series feature set. To capture the changing trend of target behavior across frames, this time series is input into a bidirectional gated recurrent unit neural network (Bi-GRU) for temporal modeling. The network input consists of N feature vectors for each frame; The Bi-GRU network consists of two layers, each with 128 units, and the activation function is the hyperbolic tangent function. The output is a 256-dimensional global scene structure encoding vector synthesized for each frame.

[0042] Ultimately, all time frame encoded vectors are combined to form a multimodal scene structure feature vector sequence, which is used for subsequent micro-motion recognition and behavior anomaly modeling.

[0043] Micro-motion modeling module: Input the multimodal scene structure feature vector into the temporal graph convolutional network to construct the inter-frame micro-motion evolution path map and extract the target's micro-change action set and abnormal behavior candidate sequence in consecutive frames.

[0044] To extract subtle motion evolution patterns of monitored targets in consecutive video frames and achieve early identification of potential abnormal behaviors, a time-dependent graph structure is constructed and graph neural network modeling is performed based on the aforementioned generated multimodal scene structure feature vectors. The specific steps include:

[0045] The multimodal scene structure feature vectors in a continuous time series are mapped to nodes in a graph structure, denoted as G=(V,E), where V is the set of nodes and E is the set of edges. Each graph node vt corresponds to a 256-dimensional feature vector generated from the image of frame t, representing the encoded result of the scene state at that time point.

[0046] The graph nodes are arranged in chronological order, allowing only edge connections with adjacent time frames. A directed edge et,t+1 is defined between each frame and its preceding and following frames, forming a linear chronological graph structure. Initially, the edges are unweighted connections used to construct the graph topology framework.

[0047] To ensure the graph structure accurately reflects the intensity of action changes and feature similarity between frames, edge weighting is necessary. A Gaussian kernel function is used to calculate the edge weights between adjacent nodes. The specific calculation method is as follows:

[0048] For any adjacent graph nodes vt and vt+1, calculate the Euclidean distance between their corresponding eigenvectors ft and ft+1, denoted as Dt; set the standard deviation parameter σ, usually set to 10; the edge weight wt,t+1 is defined using the Gaussian kernel function: When the difference in feature vectors is small, the Euclidean distance Dt is small, and the weight is close to 1; when the difference is significant, the weight approaches 0. This method is used to strengthen the representation of edges with small changes in micro-actions, weaken the influence of discontinuous behavioral edges, and form a weighted temporal graph structure.

[0049] A weighted temporal graph structure is input into a joint temporal graph modeling network combining a graph convolutional network and a gated recurrent unit to extract inter-frame state transition patterns. The specific structure is as follows: Graph Convolutional Layer Structure Definition: A two-layer Graph Convolutional Neural Network (GCN) is used, with each layer having 256-dimensional input node features and 128-dimensional output features, employing the ReLU activation function; The graph convolution operation is performed according to the following formula: the new feature of each node vt is represented by the weighted sum of the features of its neighboring nodes, where the weights are determined by the edge weights wt,t+1 and are normalized. Gated recurrent unit modeling: The graph convolution output forms a time series vector set, which is fed into a bidirectional gated recurrent unit network (Bi-GRU). The hidden state dimension of each GRU unit is 128. The bidirectional structure can capture the changing trend of the target in the direction before and after time, improving the accuracy of micro-motion evolution recognition.

[0050] After joint modeling, each time frame node outputs a set of feature vectors representing the state transition between itself and the preceding and following frames, which are called micro-action state transition features.

[0051] Based on the inter-frame state transition feature sequence output by the graph neural network, a micro-action evolution path map of the target in the temporal dimension is constructed as follows: Calculate the rate of change of features: Calculate the first-order difference for any two adjacent state feature vectors, i.e., Δft=ft+1-ft, which represents the rate of change of micro-motion from frame t to frame t+1. Determine changes in motion direction: Calculate the angle θ between Δft and Δft+1. If the angle is greater than a set threshold (e.g., 45 degrees), it indicates a sudden change in motion direction. Generate path graph structure: Connect all Δft into a vector sequence and record the rate amplitude and direction changes of each node to form a graph data structure composed of node state transitions and changes.

[0052] The final output micro-motion evolution path map can be used for subsequent tasks such as micro-motion clustering, behavior segmentation, and abnormal behavior candidate detection.

[0053] To identify subtle abnormal changes in the behavior of people or targets in surveillance videos, based on the constructed inter-frame micro-motion evolution path map, micro-motion segments of the target in consecutive frames are further extracted and potential abnormal behaviors are determined. This process includes the following steps:

[0054] In the micro-action evolution path graph, each graph node represents the multimodal scene structure feature vector of a frame image. Let the feature vector of frame t be ft, and that of frame t+1 be f(t+1).

[0055] For any two adjacent graph nodes vt and v(t+1), the element-wise difference of their corresponding feature vectors is calculated to form the inter-frame difference vector Δft. The calculation method is as follows: The sub-dimensions of this difference vector correspond to the changes in the target in the following three aspects: Spatial position change: Extract the dimensions in the feature vector that are related to the target position coordinates (such as center point, bounding box); Change in velocity direction: Extract the change value of the optical flow estimate or motion vector part; Semantic change in region: Extract the change encoding of the semantic region label to which the target belongs, such as changing from "road region" to "inactive region".

[0056] By merging these three parts to form a complete inter-frame motion difference vector sequence, a basic data structure is used to describe the evolution trend of the target's motion in a short period of time.

[0057] Based on the inter-frame motion difference vector sequence, a sliding time window clustering method is used to identify segments with similar micro-motion patterns. The specific steps are as follows:

[0058] Sliding window division: Set a fixed-length sliding time window, with a window size of 5 frames and a step size of 1 frame. That is, frames 1 to 5 are window 1, frames 2 to 6 are window 2, and so on.

[0059] Intra-window clustering: The inter-frame action difference vector within each window is used as the input sample, and K-means clustering is performed, with the initial number of clusters K set to 3. Cosine similarity is chosen as the clustering distance metric.

[0060] Similar segment extraction: For each window, if the samples inside it are mainly concentrated in a certain cluster (accounting for more than 80%), then the window is determined to have a single micro-motion pattern and is marked as a micro-motion segment.

[0061] Action set construction: Multiple consecutive micro-action segments are merged to form a micro-change action set of a target. Each action set contains a start frame, an end frame and a corresponding differential feature sequence.

[0062] This method can effectively isolate short-term consistent micro-behaviors of a target, such as slight but critical behavioral change patterns like pausing, frequent turning back, and slow walking.

[0063] For each set of micro-change actions that has been constructed, further analyze its dynamic characteristics and extract the following two types of indicators:

[0064] Rate of change curve: The magnitude of the difference vector Δft for consecutive frames within the action set is calculated to obtain the rate of change curve. The curve is used to measure the intensity of the behavior; a smooth curve indicates low dynamic behavior, while abrupt changes in the curve indicate large amplitude of movement or drastic changes in posture.

[0065] Trajectory direction offset: The target motion direction vector in each frame is normalized, and the angle θt between adjacent vectors is calculated; if θt exceeds 30 degrees, it is considered a significant direction offset. The average direction offset value is then calculated. These two indicators serve as indicators of directional changes in the overall behavior. As core descriptive quantities for behavioral assessment, they provide criteria for subsequent comparisons with normal behavioral patterns.

[0066] Each set of subtle changes in actions is compared with a pre-set historical normal behavior template in the system to perform deviation calculation and anomaly identification. Specifically:

[0067] Historical template construction method: In the early stage of system deployment or training phase, based on large sample monitoring data, collect typical normal behavior segments of the target, such as normal walking, staying, obstacle avoidance, etc.; construct a template library through the same feature extraction method, and generate 1 to 3 standard template feature sequences for each type of behavior.

[0068] Deviation calculation: Dynamic time warping is performed on the feature sequence of the action set to be analyzed and the sequence of each type of behavior template to calculate the minimum alignment distance. Let the minimum distance be dmin. If dmin exceeds the threshold set for that type of template (e.g., 1.5 times the average training distance), then the action set is considered to deviate from normal behavior.

[0069] Anomaly candidate labeling: For action sets whose deviation exceeds the threshold, record the corresponding time period and target ID, and label them as anomaly behavior candidate sequences; this sequence will be sent to the subsequent anomaly scoring and response module for further classification and response strategy triggering.

[0070] Frequency map construction module: Based on the inter-frame micro-motion evolution path map, calculate the micro-motion co-occurrence frequency map of each target within the historical time window, and construct a multi-target behavior coupling map.

[0071] To characterize the potential behavioral relationships between different monitoring targets in a park setting, a quantifiable and modelable multi-target behavior coupling map is constructed based on the inter-frame micro-motion evolution path map and micro-change action set information. The specific steps include:

[0072] Within a set historical time window (e.g., 30 seconds corresponding to 750 frames), the micro-change motion sets of all targets are traversed. The motion set of each target includes a start frame, an end frame, and its corresponding differential motion vector sequence.

[0073] For any two targets A and B, if their action sets overlap in time, meaning the time intervals of the two action segments overlap, this is considered a "micro-action co-occurrence event." The definition is as follows:

[0074] Divide the timeline into non-overlapping sub-windows (e.g., one sub-window for every 5 frames).

[0075] If two targets each have at least one valid motion segment within the same sub-window (the inter-frame change amplitude is greater than the set value of 0.3), then the co-occurrence count is incremented by 1;

[0076] The number of times target A and target B co-occur within the entire historical time window is denoted as NAB.

[0077] The number of co-occurrences serves as a preliminary quantitative basis for the potential behavioral associations between targets.

[0078] To comprehensively consider both the spatial proximity and behavioral synchronization between targets, the change in spatial distance needs to be introduced as a weighting correction factor. The specific calculation is as follows:

[0079] Spatial distance change calculation: For each micro-motion co-occurrence event, extract the two-dimensional image coordinates (such as center point coordinates) of targets A and B in that frame; calculate the Euclidean distance dAB, and average the distance values ​​of all co-occurrence events to obtain the average spatial distance DAB. Normalization processing: For all target pairs, normalize the co-occurrence frequency NAB and the average spatial distance DAB to the [0,1] interval, using the maximum-minimum normalization formula: Normalized co-occurrence frequency is: Wfreq(A,B)=(NAB-Nmin)÷(Nmax-Nmin); Normalized distance weight is: Wdist(A,B)=1-(DAB-Dmin)÷(Dmax-Dmin); Weight fusion: Linearly weighted fusion of the two normalization indices forms the final co-occurrence frequency weight, as follows: WAB=0.6×Wfreq(A,B)+0.4×Wdist(A,B).

[0080] Through the above calculations, a symmetric inter-target correlation weight matrix W is obtained, where each element WAB represents the coupling strength of the micro-actions between targets A and B within the time window.

[0081] Based on the weight matrix W above, the behavioral relationship graph structure between multiple objectives is constructed as follows:

[0082] Set an edge connection threshold: Compare the co-occurrence frequency weight WAB with the set threshold Tconn (e.g., 0.65); if WAB ≥ Tconn, then target A and target B are considered to have a significant behavioral association within the current window, and a connection edge eAB is established.

[0083] Constructing a graph structure: All targets are treated as nodes in the graph, and each edge satisfying the connection condition is an undirected weighted edge, with the edge weight being WAB; resulting in an undirected graph Gassoc=(V,E), where V is the set of targets and E is the set of behavioral coupling edges. This graph structure reflects the co-occurrence relationship and spatial proximity of micro-actions among targets within the same time period, possessing high-resolution behavioral linkage expression capabilities.

[0084] The multi-objective behavior association graph structure is expanded along the time dimension, and the graph structure is repeatedly constructed within different time windows to form a time series graph set. Furthermore, the connection strength and duration between objectives are analyzed to extract stable behavioral coupling relationships, as follows:

[0085] Calculate connection persistence: For target A and target B, whether edge eAB exists continuously in multiple consecutive time windows; if the number of consecutive occurrences exceeds 60% of the window number, it is determined to be a stable behavioral coupling relationship.

[0086] Graph construction: Mark all stable coupling relationships in the graph as "long-term edges" and attach edge attributes such as frequency of occurrence, average weight, and duration;

[0087] Ultimately, a multi-target behavior coupling graph with time tags, behavioral relationships, and target identity information is formed, which is used for subsequent abnormal relationship identification and multi-target behavior scoring.

[0088] Abnormal Behavior Analysis Module: Based on the multi-target behavior coupling map, construct an abnormal behavior scoring model, perform risk scoring and clustering classification on abnormal interaction behaviors between targets, and identify potential security risk behaviors.

[0089] To identify target pairs with potential threats or abnormal behavior in the park and improve the accuracy of multi-target interaction behavior analysis, this study models and analyzes the characteristics of target interaction edges based on the constructed multi-target behavior coupling graph. A risk level is output through an anomaly scoring model, followed by clustering classification and high-risk behavior labeling. The specific steps are as follows:

[0090] Based on the connection edges between any two target nodes in the multi-target behavior coupling graph, multi-dimensional features describing their interaction behavior states are extracted. Each connection edge Eij (connecting target i and target j) is considered a behavior sample, and its features include:

[0091] Temporal continuity characteristic: This involves calculating the proportion of times the connecting edge appears within a continuous time window. Dividing the historical time period into N time windows, if the target pair remains connected for M of these windows, the temporal continuity value is M ÷ N.

[0092] Co-occurrence frequency weight: The value of Wij in the aforementioned correlation weight matrix is ​​directly extracted to represent the comprehensive intensity of the co-occurrence frequency of micro-actions between the two targets.

[0093] Spatial proximity: Calculate the average Euclidean distance between target i and target j in all historical co-occurrence frames, denoted as Davg, and then use max-min normalization to map to the interval [0,1]. The smaller the value, the closer they are.

[0094] The above three sub-features are concatenated into a set of three-dimensional behavioral interaction feature vectors in a unified format, Fij=[Tcont,Wij,Dnorm], which are used to describe the interactive behavioral features of the target pair.

[0095] A supervised learning method is used to construct an anomaly scoring model, and behavioral risk is estimated from the above interaction feature vectors. The specific steps are as follows:

[0096] Training sample preparation: Using historical monitoring data and manual annotation results, typical interactive behavior segments of three categories, namely "normal", "minor abnormal" and "serious abnormal", were collected, and the corresponding interactive feature vectors and risk level labels were extracted to construct a training sample set.

[0097] Model Selection and Training: The Gradient Boosting Decision Tree algorithm is adopted as the anomaly scoring model, which has the advantages of strong ability to identify small samples and nonlinear boundaries. The model input is a three-dimensional interactive feature vector, and the output is a risk score R with a continuous value range [0,1]. The higher the score, the more abnormal the behavior.

[0098] Model output explanation: After training, input the behavioral feature vector Fij of any target pair, and the model outputs its risk score Rij, which is used for subsequent cluster analysis and high-risk identification.

[0099] To achieve unsupervised risk level classification, clustering based on scoring results is performed on all interaction samples. The steps are as follows:

[0100] Clustering algorithm selection: Density-Based Spatial Clustering of Applications with Noise (DBSCAN) was used to cluster the scoring results. This algorithm can identify clusters of arbitrary shapes and effectively remove outliers.

[0101] Parameter settings: The minimum number of samples is set to 10; the neighborhood radius ε is set to 0.1, meaning that samples with a score difference within 0.1 are considered neighbors.

[0102] Risk classification results: After clustering, all interaction samples were divided into three categories: those clustered in [0, 0.3] were low-risk behaviors; those clustered in [0.3, 0.6] were suspicious behaviors; and those clustered in [0.6, 1.0] were high-risk behaviors. The clustering results provide a basis for behavioral levels for subsequent action responses.

[0103] Further analysis was performed on target pairs classified as high-risk behaviors in the clustering results, and behavioral semantic recognition was performed by combining their time trajectories and semantic regions:

[0104] Trajectory analysis: Analyze the position sequence of high-risk targets within the historical window. If abnormal trajectory behavior (such as repeated approach, detour, or parallel movement) is found, the risk level is further increased.

[0105] Semantic region cross-judgment: The frame location where the high-risk target behavior occurs is mapped to the static semantic region map. If it occurs in a highly sensitive area (such as a blind spot in the passage, a dangerous goods storage point, or an exit intersection), the behavior is marked as a potential security risk behavior.

[0106] The final output is structured labeled data consisting of "target number + time period + behavior level + area location", which can be used for subsequent monitoring and linkage response or manual intervention.

[0107] To verify the effectiveness of the behavior anomaly scoring model based on multi-objective behavior coupling graph described in this invention in a real-world park monitoring environment, a comparative experiment was conducted using the following methods: Experimental environment and dataset settings: Monitoring area: A high-security industrial park, with a total area of ​​approximately 15,000 square meters; Number of cameras deployed: 24, covering the main passage, entrances and exits, parking lot, and chemical storage area; Experiment duration: 72 hours of continuous video data acquisition; Data annotation method: The interaction behaviors were annotated frame by frame by 5 security experts, and a total of 674 behavioral events were annotated, including 65 high-risk events, 104 suspicious behaviors, and the rest were normal behaviors; Comparison method: Compare with traditional single-target trajectory anomaly detection methods (such as trajectory deviation detection + basic behavior template comparison).

[0108] Key Indicator Comparison:

[0109] The results are analyzed as follows: Significantly improved recognition accuracy: This method effectively captures abnormal group behaviors, such as coordinated loitering, linked contact, and abnormal approach, by introducing multi-target coupling relationships and micro-movement co-occurrence modeling, significantly improving the accuracy of high-risk behavior recognition (by about 14 percentage points). The false alarm rate is significantly reduced: thanks to the scoring clustering mechanism and spatial semantic constraints, it has a higher fault tolerance for non-abnormal behaviors such as "crossing paths" and "temporary approach", and the false alarm rate is reduced by more than 60%; More timely response: The graph structure and feature encoding run in parallel, combined with graph neural networks and lightweight clustering models, to achieve identification within an average of 1.6 seconds, meeting the needs of real-time early warning; Highly scalable: This scoring model supports continuous learning and updates, and can be quickly migrated and deployed in different park scenarios by adding labeled samples.

[0110] Intelligent linkage acquisition module: Based on the risk score results, it automatically generates a coordinate system of microscopic observation features in high-risk areas, and controls the camera to adjust its angle and focal length, performing high-frequency image acquisition and local fine-grained recognition and analysis on key targets.

[0111] To improve the response accuracy and image recognition quality of high-risk behavioral events in the park, based on the aforementioned abnormal behavior scoring results, microscopic-level image acquisition and recognition analysis are automatically performed in high-risk areas for key targets with scores exceeding the threshold.

[0112] When the behavioral anomaly scoring model outputs a risk score value for a target that exceeds a preset threshold (e.g., 0.65), and the target is located in a highly sensitive semantic area (e.g., entrance / exit, closed lane, chemical warehouse, etc.), the micro-observation process is immediately triggered.

[0113] First, a two-dimensional feature coordinate system is constructed with the current position of the target as the origin. The specific steps are as follows: Determine the origin coordinates: Extract the image coordinates (x, y) of the target in the current frame from the multimodal scene structure feature vector and convert them into the corresponding world coordinates (X, Y); Add a behavior direction axis: Based on the spatial trajectory of the target in the past 5 consecutive frames, fit the trajectory direction vector Vdir using the least squares method, and set this direction as the X-axis; Establish regional boundary constraints: Combine semantic labels in the region segmentation map to mark the visible region boundary of the target's region, which serves as the effective activity range within the coordinate system.

[0114] Ultimately, a microscopic observation feature coordinate system is formed, which includes a spatial origin, direction of motion trend, semantic boundary, and visual constraints, to guide the next step of camera selection and attitude adjustment.

[0115] Based on the target's motion trend and spatial position in the microscopic observation feature coordinate system, the camera group with the best coverage capability is selected, as follows: Camera field of view model construction: The field of view coverage of each camera is preset to be a fan-shaped area, including the starting angle, the ending angle, and the farthest viewing distance corresponding to the maximum focal length. The spatial imaging area can be calculated by combining the installation position and the pitch angle. Viewpoint accessibility judgment: For each candidate camera, calculate the angle θ between its current main viewing direction and the target direction vector. If θ ≤ 45 degrees and the target is within its effective imaging range, it is determined as an accessible viewpoint; Preferred camera group: For all cameras with accessible viewpoints, construct a multi-factor scoring function based on the spatial distance from the target, the cosine value of the angle, and the historical imaging clarity score, and select the top 2 cameras with the highest scores to enter the linkage control queue.

[0116] Perform real-time linkage control on the preferred camera group, adjust its angle and focal length parameters to ensure that the shooting area accurately covers the path prediction range of the target: Path prediction method: Use the Kalman filter algorithm to predict the trajectory of the target in the next 10 frames, obtain the potential moving path area of the target, and output it as a rectangular spatial area R; Angle adjustment control: Calculate the offset angle between the current shooting direction of the camera and the center of area R, and control the pitch motor and the horizontal rotation motor to adjust the angle in a proportional manner, with the adjustment accuracy set to 1 degree; Focal length adaptive adjustment: Based on the size of the target in the image and the imaging clarity evaluation value (using the structural similarity index SSIM, and setting the target imaging SSIM ≥ 0.75 as qualified), adjust the parameters of the zoom motor in real time to make the target image occupy between 30% and 50% of the height of the image.

[0117] The above control process is executed at a refresh frequency of 10 times per second to ensure the continuous capture accuracy of the target during movement.

[0118] When the target stably enters the central area of the microscopic observation area (within a radius of 1 meter from the origin of the coordinate system), the system immediately executes high-frequency image acquisition and multi-channel recognition analysis: Image acquisition frequency setting: The acquisition frequency is set to 10 frames per second, the continuous acquisition duration is 5 seconds, a total of 50 frames of images are acquired, and they are cached to the high-risk buffer; Local area extraction: Extract the target bounding box area in each frame of the image and perform the following recognition tasks: Face recognition: Use the multi-scale face detector MTCNN model to extract the face area and output the feature vector; Clothing texture recognition: Identify the texture patterns of the upper and lower clothes through the color histogram and the texture direction encoding sub-model; Carried object detection: Use the YOLOv5 model to perform object recognition on the surrounding area of the target, and identify the types and location information of objects such as backpacks and cardboard boxes.

[0119] Encoding and storage of recognition results: Encode the above recognition results into feature vectors, bind them to the original image frames, and store them in the event database for subsequent behavior confirmation, traceability, and linkage response.

[0120] Response module: Compare the analysis result with the preset security behavior template library, output the evaluation result of the abnormal behavior level, and trigger the response strategy.

[0121] To achieve the automatic recognition and rapid response to high-risk target behaviors, based on the behavior recognition results and atlas information extracted in the foregoing steps of the present invention, compare them with the preset security behavior template library for analysis, judge their abnormal levels, and trigger different-level linkage response strategies accordingly. This process specifically includes the following: The security behavior template library is established based on historical behavior data and expert experience rules in typical scenarios of the park, and the content includes: Template classification dimensions: The templates are classified according to three dimensions: behavior type, occurrence area, duration, and movement mode, such as "normal passage template", "short stay template", "illegal approach template"; Feature structure definition: Each template contains the following structure: Behavior trajectory pattern (such as path direction, speed change); Semantic area distribution (sequence of area labels where the behavior occurs); Micro-action feature distribution (statistical features of differential vector sequences, such as average modulus length and direction change rate); Coding format: All template behaviors are encoded in the form of unified multi-dimensional feature vectors and stored in the local template database to support fast searching and matching.

[0122] Compare the target behavior segment to be evaluated (output by the scoring module or the fine-grained recognition module) with the security behavior template library one by one, and the steps are as follows: Behavior feature encoding: Extract the features of the trajectory, semantic distribution, and micro-action sequence of the target behavior within the analysis time window, and encode them into a behavior vector Btest, which has the same structure as the template behavior vector; Similarity calculation method: Use the Dynamic Time Warping algorithm to perform matching operations on Btest and each template vector, and output the minimum matching distance Dmin; Determination of the nearest neighbor template: If Dmin is lower than the similarity threshold Tcat corresponding to the template category (for example, set to 0.15 for "normal passage" and 0.3 for "illegal approach"), it is considered that the target behavior conforms to this template, and the matching label is output. If the matching degree of Btest with any template is lower than the lower limit of the threshold of all categories, it is determined as an undefined abnormal behavior.

[0123] Combine the matching result with the original score value to conduct an abnormal level evaluation, and the rules are as follows: Score threshold setting: Divide the risk score result R into three levels: Low risk: 0 ≤ R < 0.4; Medium risk: 0.4 ≤ R < 0.7; High risk: R ≥ 0.7; Template type determination weight: If the matched template belongs to a highly sensitive behavior type (such as "abnormal return" or "illegal gathering"), its risk level will be automatically increased by one level.

[0124] The final evaluation results are output in a structured format as “Target ID + Behavior Time Period + Anomaly Level + Matching Template Name” for the response module to use.

[0125] Based on the level of abnormal behavior, the following response process will be automatically executed: Low-risk behavior response strategy: Record event information, mark and archive it as training samples for subsequent data, without triggering real-time intervention; Medium-risk behavior response strategy: Trigger the front-end alert mechanism, such as popping up a prompt box on the monitoring platform and linking it to a voice broadcast warning; High-risk behavior response strategy: Link up with cameras for continuous tracking and data collection, and push the target ID and image information to the security control center to achieve real-time early warning; If the target location is close to a restricted area, the park's access control system, lighting equipment, or warning lights will be used in conjunction to implement a preset blockade strategy. It can automatically retrieve the target's historical behavior trajectory for retrospective analysis and comparative tracking.

[0126] All response strategies are configurable and can be customized and have permissions set according to the actual scenario of the park.

[0127] This invention effectively improves the semantic accuracy and real-time response of abnormal behavior identification by introducing behavior template comparison and multi-level response logic.

[0128] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A computer vision-based intelligent monitoring system for industrial parks, characterized in that: include: Video acquisition module: Collects video image frame sequences from multiple surveillance cameras within the park at different time periods, and constructs a video stream map that is spatially synchronized across cameras; Feature extraction module: Based on the video stream graph, a region segmentation algorithm is used to extract static structural features and dynamic change features from each image frame to generate a multimodal scene structural feature vector; Micro-motion modeling module: Input the multimodal scene structure feature vector into the temporal graph convolutional network to construct an inter-frame micro-motion evolution path map and extract the target's micro-change action set and abnormal behavior candidate sequence in consecutive frames; Frequency map construction module: Based on the inter-frame micro-motion evolution path map, calculate the co-occurrence frequency map of micro-motions of each target within the historical time window, and construct a multi-target behavior coupling map; Abnormal Behavior Analysis Module: Based on the multi-target behavior coupling map, construct an abnormal behavior scoring model, perform risk scoring and clustering classification on abnormal interaction behaviors between targets, and identify potential security risk behaviors; Intelligent linkage acquisition module: Based on the risk score results, it automatically generates a microscopic observation feature coordinate system in high-risk areas, and controls the camera to adjust its angle and focal length, performing high-frequency image acquisition and local fine-grained recognition analysis on key targets; Response module: Compares the analysis results with the preset security behavior template library, outputs the abnormal behavior level assessment results, and triggers the response strategy.

2. The intelligent campus monitoring system based on computer vision according to claim 1, characterized in that: The process of constructing a cross-camera spatially synchronized video stream graph includes the following steps: The system collects a sequence of video image frames continuously captured by fixed-view surveillance cameras deployed in different areas of the park within a preset time period; timestamps are aligned between different video image frames based on the spatial location information and time synchronization signals of each camera; target correspondence is calculated for image frames within the coverage area of ​​adjacent cameras using a feature point matching algorithm, and a target trajectory association model is constructed between cameras; a video stream map under unified spatial coordinates is generated based on the continuity of the target trajectory and scene structure information.

3. The intelligent campus monitoring system based on computer vision according to claim 1, characterized in that: The process of extracting static structural features and dynamic change features from each image frame using a region segmentation algorithm includes the following steps: The image frames in the video stream atlas are processed for region segmentation. A pre-trained semantic segmentation model is used to classify road regions, building structure regions and activity regions at the pixel level to generate a static structural region distribution map. Based on the static structural region distribution map, inter-frame difference calculation and optical flow field estimation are performed on continuous image frames to extract pixel motion change information of the target in the active region and generate dynamic change feature map. Based on the dynamic change feature map and the static structure region distribution map, the target region is subjected to feature weighted fusion calculation to form a scene structure feature set of unified dimension; The scene structure feature set is vectorized and encoded in chronological order to generate a multimodal scene structure feature vector.

4. The intelligent campus monitoring system based on computer vision according to claim 1, characterized in that: The construction of the inter-frame micro-motion evolution path map includes the following steps: The multimodal scene structure feature vectors are used to construct a node graph structure in chronological order, where each frame of the image corresponds to a graph node, and initial edge connections are established between nodes based on their chronological order. Based on the feature similarity and time interval between nodes, a Gaussian kernel function is used to calculate the edge weights and construct a weighted time sequence graph structure. The weighted temporal graph is input into a joint temporal graph convolutional network containing graph convolutional layers and gated recurrent units to extract the micro-action state transition features between nodes; Based on the difference in feature change rate and action vector direction between nodes, an inter-frame micro-motion evolution path map is generated.

5. The intelligent campus monitoring system based on computer vision according to claim 4, characterized in that: The extraction of the target's micro-change action set and abnormal behavior candidate sequence in consecutive frames includes the following steps: The difference is calculated on the target feature vectors corresponding to adjacent graph nodes in the inter-frame micro-motion evolution path map. The changes in the target's spatial position, velocity direction and regional semantics at each moment are extracted to form an inter-frame motion difference vector. Based on the inter-frame motion difference vector, a sliding time window clustering algorithm is used to identify micro-motion segments with similar change patterns in a continuous frame sequence, and to construct a set of micro-change motions of the target. For each micro-motion segment, calculate its rate of change curve and trajectory direction offset value, and compare the deviation with the historical normal behavior template; if the deviation exceeds the set threshold, mark the corresponding frame sequence as an abnormal behavior candidate sequence.

6. The intelligent campus monitoring system based on computer vision according to claim 1, characterized in that: The construction of a multi-objective behavior coupling map includes the following steps: Based on the inter-frame micro-motion evolution path map, the micro-change action set of each target in the continuous time series is extracted, and the time overlap interval of different target micro-motion segments is counted within a preset historical time window to calculate the number of micro-motion co-occurrences between targets. Based on the co-occurrence frequency and the change in spatial distance between targets, normalization is used to calculate the co-occurrence frequency weight of micro-actions between targets, and a correlation weight matrix between targets is constructed. Based on the association weight matrix, targets with co-occurrence frequencies higher than a set threshold are connected to generate a multi-target behavior association graph structure. Based on the connection strength and temporal continuity between each objective, a multi-objective behavior coupling map is formed.

7. The computer vision-based intelligent monitoring system for industrial parks according to claim 6, characterized in that: The risk scoring and clustering classification of abnormal interactions between targets includes the following steps: Extract the temporal continuity, co-occurrence frequency weights, and spatial proximity features of each connection edge in the multi-objective behavior coupling graph, and construct a multi-dimensional behavior interaction feature vector. The feature vector is input into the behavior anomaly scoring model trained by supervised learning, and the risk score value of each group of target interaction behaviors is output. Based on the risk score, density clustering algorithm is used to cluster and classify behavioral samples, dividing them into low-risk, suspicious and high-risk behavior categories; Target pairs in high-risk categories are identified and, combined with their spatiotemporal trajectories and semantic regions, are marked as potential security risk behaviors.

8. The intelligent campus monitoring system based on computer vision according to claim 7, characterized in that: The linkage control of the camera to adjust its angle and focus includes the following steps: Based on the location of high-risk targets and their semantic regions output by the abnormal behavior scoring model, a microscopic observation feature coordinate system is constructed, which includes information on target location, movement trend, and surrounding structure. Based on the target's velocity direction vector and regional occlusion conditions in the coordinate system, a camera group with the best field of view coverage is selected. The selected cameras are controlled to adjust their pitch angle, horizontal rotation angle, and focal length parameters in real time so that their main field of view is aligned with the target's movement path.

Images