Video monitoring abnormal behavior identification method and system based on edge AI

Abstract

The invention discloses a video monitoring abnormal behavior identification method and system based on edge AI, and relates to the technical field of intelligent video analysis, and the method comprises the steps: carrying out the frame segmentation processing of a video monitoring data stream, obtaining a video frame sequence, building a target trajectory prediction model, predicting the position region of a current frame, and generating target position prediction data; performing multi-scale feature extraction on the video frame sequence, performing fusion matching on a target feature vector and target position prediction data, and constructing an enhanced feature matrix; carrying out weight distribution on the key behavior characteristics by adopting an attention mechanism, setting a dynamic threshold adjustment mechanism, and dynamically adjusting an abnormal behavior judgment threshold according to the personnel density; and inputting the adjusted feature data into an abnormal behavior classifier for identification and judgment, outputting an abnormal behavior identification result, and generating an abnormal event report. According to the method, the abnormal behavior detection accuracy in a complex monitoring scene is improved, the false report and missing report rate is reduced, and the millisecond-level real-time response capability is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent video analysis technology, and in particular to a method and system for identifying abnormal behavior in video surveillance based on edge AI. Background Technology

[0002] Traditional methods primarily rely on static feature extraction and rule matching, such as optical flow analysis and spatiotemporal interest point detection. However, limited by computational complexity and generalization ability, they struggle to meet real-time monitoring requirements. In recent years, the introduction of deep learning technology has significantly improved the accuracy of behavior recognition, while the rise of edge AI has further driven the development of low-latency, high-efficiency anomaly detection solutions. However, existing technologies still face challenges such as insufficient target tracking accuracy, low feature fusion efficiency, and poor adaptability to dynamic scenes. In particular, the high false positive and false negative rates in complex environments limit their practical application in edge computing environments.

[0003] CN119904785B proposes an anomaly recognition method based on video compression and keyframe extraction, which improves processing efficiency by reducing the amount of data. Although this method reduces storage and computational burden, it may lead to the loss of key behavioral information due to its reliance on a fixed keyframe extraction strategy, and it does not consider the continuity of the target's motion trajectory, resulting in insufficient robustness in dynamic scenes.

[0004] CN108171214A employs a combination of inter-frame difference comparison and deep learning to improve anomaly detection accuracy. However, it relies on static thresholds for anomaly judgment, making it difficult to adapt to changes in different scenarios (such as sparse / dense crowds), and it does not fully utilize spatiotemporal feature correlations, resulting in a high false positive rate in complex environments. Furthermore, neither of these methods is optimized for edge computing environments, making it difficult to achieve efficient real-time analysis on resource-constrained devices.

[0005] In summary, existing video surveillance anomaly behavior recognition technologies still suffer from the following key problems: insufficient target trajectory prediction accuracy, leading to tracking failure in occluded or fast-moving scenarios; simplistic feature fusion strategies that fail to fully exploit the correlation between spatiotemporal features; and anomaly detection relying on fixed thresholds and lacking dynamic adjustment mechanisms, thus affecting the system's generalization ability. To address these issues, this paper proposes an anomaly behavior recognition method and system based on multi-scale feature fusion and dynamic threshold adjustment using edge AI. Summary of the Invention

[0006] In view of the problems of insufficient target trajectory prediction accuracy, low feature fusion efficiency and poor adaptability to dynamic scenes in existing abnormal behavior recognition technologies, this invention is proposed.

[0007] Therefore, the problem to be solved by this invention is how to achieve high-precision, low-latency edge-end abnormal behavior recognition, while improving detection robustness and adaptability in complex scenarios.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] In a first aspect, embodiments of the present invention provide a method for identifying abnormal behavior in video surveillance based on edge AI, comprising,

[0010] The video surveillance data stream is segmented to obtain a video frame sequence. At the same time, a target trajectory prediction model is established to predict the location region of the current frame based on the target location information, and target location prediction data is generated.

[0011] Multi-scale feature extraction is performed on the video frame sequence to obtain the target feature vector, and the target feature vector is fused and matched with the target position prediction data to construct an enhanced feature matrix;

[0012] Based on the enhanced feature matrix, an attention mechanism is used to assign weights to key behavioral features, and a dynamic threshold adjustment mechanism is set to dynamically adjust the abnormal behavior judgment threshold according to the personnel density of the current scene.

[0013] The adjusted feature data is input into the abnormal behavior classifier for identification and judgment, and the abnormal behavior identification results are output to generate an abnormal event report.

[0014] As a preferred embodiment of the video surveillance abnormal behavior recognition method based on edge AI described in this invention, the method includes: inputting adjusted feature data into an abnormal behavior classifier for recognition and judgment, outputting abnormal behavior recognition results, and generating an abnormal event report, including:

[0015] Construct an abnormal behavior classifier, wherein the abnormal behavior classifier adopts a depthwise separable convolutional structure, including a feature encoding layer, a behavior pattern recognition layer, and an anomaly determination layer;

[0016] The feature data adjusted by dynamic thresholds are grouped and organized according to the target identity to construct a target feature dataset. The target feature dataset is then segmented temporally to generate fixed-length feature sequence fragments.

[0017] The feature sequence fragment is input into the feature encoding layer, and a one-dimensional convolution kernel is used to perform a convolution operation on the feature sequence fragment to extract local feature patterns.

[0018] The encoded feature vector is transmitted to the behavior pattern recognition layer, and a bidirectional long short-term memory network is used to perform temporal modeling on the encoded feature vector to capture the target's behavioral continuity features and motion trend features.

[0019] The behavior pattern recognition layer outputs a behavior feature representation vector, and the behavior feature representation vector is compared with the abnormal behavior determination threshold.

[0020] The behavior feature representation vector that has been filtered by the threshold is input into the anomaly determination layer, wherein the anomaly determination layer includes a classification head for several abnormal behavior categories;

[0021] A fully connected layer is used to perform a linear transformation on the behavior feature representation vector to generate an abnormal behavior probability distribution vector.

[0022] The probability distribution vector of the abnormal behavior is subjected to Softmax normalization to obtain the confidence score of each abnormal behavior category, and the abnormal behavior category with the first confidence score is selected as the recognition result.

[0023] If an abnormal behavior event is detected, the spatiotemporal location information of the abnormal target is extracted from the residual feature matrix R, and the spatiotemporal location information is associated and bound with the abnormal behavior identification result to generate an abnormal behavior event record.

[0024] Based on the abnormal behavior event records, an abnormal event report is constructed, and the abnormal event report is sent to the monitoring center or mobile terminal through the communication module of the edge AI device to complete the rapid detection and alarm process of abnormal behavior.

[0025] As a preferred embodiment of the video surveillance abnormal behavior recognition method based on edge AI described in this invention, the recognition result includes:

[0026] A multi-dimensional judgment mechanism is established based on the correlation between the first confidence score and the density level threshold.

[0027] When the personnel density value is at the first-level density level and the first confidence score exceeds the first-level threshold, the entropy value of the abnormal behavior probability distribution vector is calculated at the same time; if the entropy value is lower than the preset entropy value threshold, the abnormal behavior identification result is confirmed as a high-confidence abnormal event.

[0028] When the personnel density value is at the second-level density level and the first confidence score is between the second-level threshold and the first-level threshold, the temporal change gradient of the behavior feature representation vector is extracted; if the temporal change gradient exceeds the preset gradient threshold, the abnormal behavior identification result is marked as a gradual abnormal event.

[0029] When the difference between the first confidence score and the second confidence score is less than the preset confidence competition threshold, the abnormal behavior category competition judgment is triggered.

[0030] As a preferred embodiment of the video surveillance abnormal behavior recognition method based on edge AI described in this invention, the abnormal behavior category competition judgment includes:

[0031] If the population density value is at level three and the first confidence score corresponds to gathering behavior and the second confidence score corresponds to escape behavior, then it is determined to be an abnormal crowd evacuation event.

[0032] If the population density value is at level four and the first confidence score corresponds to fighting behavior and the second confidence score corresponds to gathering behavior, then it is determined to be an abnormal group conflict event.

[0033] When the confidence threshold condition is met and the personnel density value is at the first density level, the abnormal behavior category corresponding to the first confidence score is selected as the identification result.

[0034] As a preferred embodiment of the video surveillance abnormal behavior recognition method based on edge AI described in this invention, the method includes: based on the enhanced feature matrix, using an attention mechanism to assign weights to key behavioral features, and setting a dynamic threshold adjustment mechanism to dynamically adjust the abnormal behavior judgment threshold according to the personnel density of the current scene, including:

[0035] The enhanced feature matrix is ​​transformed through a linear transformation layer to generate the query matrix Q, the key matrix K, and the value matrix V, respectively.

[0036] Calculate the matrix product of the query matrix Q and the transpose of the key matrix K to obtain the attention score matrix S;

[0037] Divide the elements of the attention score matrix S by the scaling factor to obtain the scaled attention score matrix;

[0038] The rows of the scaled attention score matrix are subjected to Softmax normalization to generate the attention weight matrix A;

[0039] Perform a matrix multiplication operation between the attention weight matrix A and the value matrix V to obtain the weighted feature matrix W;

[0040] Perform a residual concatenation operation between the weighted feature matrix W and the enhanced feature matrix to generate a residual feature matrix R;

[0041] Based on the target location prediction data, calculate the number of people in the current video frame and the area of ​​the monitored area to obtain the people count and people density values.

[0042] Establish a dynamic mapping function between personnel density and abnormal behavior judgment threshold, divide the level range according to the personnel density value, and set a corresponding threshold adjustment coefficient for the density level;

[0043] Set a baseline abnormal behavior judgment threshold, input the personnel density value into the dynamic mapping function, obtain the threshold adjustment coefficient corresponding to the current scene, and generate the abnormal behavior judgment threshold through numerical calculation;

[0044] The residual feature matrix R and the abnormal behavior judgment threshold are correlated, and the feature weights of the residual feature matrix R are adjusted a second time to generate feature data that has been dynamically thresholded, which is then used as the input to the abnormal behavior classifier.

[0045] As a preferred embodiment of the video surveillance abnormal behavior recognition method based on edge AI described in this invention, the method for establishing the enhanced feature matrix is ​​as follows:

[0046] Convolutional kernel groups are constructed, and the number of channels in the convolutional kernel groups is configured to increase incrementally to form a multi-scale feature extraction network architecture, wherein the convolutional kernel groups include small-sized convolutional kernels, medium-sized convolutional kernels, and large-sized convolutional kernels;

[0047] The pixel values ​​of the video frame sequence are normalized to map the range of pixel values ​​to a standard numerical range, thereby generating standardized video frame sequence data.

[0048] The standardized video frame sequence data is input into the multi-scale feature extraction network architecture, and features are extracted through the convolutional kernel group to obtain fine-grained feature maps, medium-grained feature maps, and coarse-grained feature maps;

[0049] The fine-grained feature map, medium-grained feature map, and coarse-grained feature map are concatenated along the channel dimension to generate a multi-scale fused feature map.

[0050] The multi-scale fused feature map is reduced in dimensionality by using a dimensionality-reducing convolution kernel to compress the number of feature channels to a preset dimension, thereby obtaining a compressed multi-scale feature map.

[0051] Based on the predicted center coordinates and predicted region boundary information of the target location prediction data, the spatial region where the target is located is located on the compressed multi-scale feature map, and the RoI pooling algorithm is used to extract the feature sub-map corresponding to the target region.

[0052] The prediction center coordinates, prediction area boundaries, and prediction confidence information of the target location prediction data are numerically encoded to generate a location encoding vector.

[0053] The target feature vector and the position encoding vector are concatenated to generate a high-dimensional fusion feature vector, which is used as the enhanced feature vector.

[0054] Based on the video frame sequence of personnel and object targets, the corresponding enhanced feature vectors are obtained, and the enhanced feature vectors of all targets are arranged according to the target identity to construct an enhanced feature matrix.

[0055] As a preferred embodiment of the video surveillance abnormal behavior recognition method based on edge AI described in this invention, the method includes: performing frame segmentation processing on the video surveillance data stream to obtain a video frame sequence; simultaneously establishing a target trajectory prediction model; predicting the location region of the current frame based on target location information; and generating target location prediction data, including:

[0056] The system acquires real-time video surveillance data streams from video surveillance equipment, performs format conversion and preprocessing on the video surveillance data streams, and obtains preprocessed video surveillance data streams.

[0057] The preprocessed video surveillance data stream is divided into time-axis segments according to a preset frame rate, and the video surveillance data stream is decomposed into discrete single-frame images to construct a video frame sequence.

[0058] The video frames of the video frame sequence are subjected to target detection processing to identify personnel targets and object targets in the video frames, and the bounding box localization algorithm is used to determine the position coordinates of each target in the current video frame to generate a target position information dataset.

[0059] Based on the target location information dataset, a target tracking algorithm is used to perform identity association matching on the same target in adjacent video frames to construct target trajectory data, wherein the target trajectory data includes target identity identifiers and historical location sequences;

[0060] Based on the target trajectory data, a target trajectory prediction model is established, wherein the target trajectory prediction model adopts the Kalman filter algorithm;

[0061] Extract target location information from N consecutive frames of the target trajectory data to form a target historical location sequence input vector;

[0062] The target historical location sequence input vector is input into the target trajectory prediction model to calculate and predict the target's location region in the current frame, and generate a location prediction result.

[0063] Based on the location prediction results, target location prediction data is generated, and the target location prediction data is bound with the corresponding target identity identifier to form a target location prediction dataset.

[0064] Secondly, this invention provides a video surveillance abnormal behavior recognition system based on edge AI, comprising:

[0065] The frame processing and trajectory prediction module is used to perform frame segmentation processing on the video surveillance data stream to obtain the video frame sequence. At the same time, it establishes a target trajectory prediction model, predicts the location region of the current frame based on the target location information, and generates target location prediction data.

[0066] The feature extraction and fusion module is used to extract multi-scale features from the video frame sequence to obtain target feature vectors, and to fuse and match the target feature vectors with the target position prediction data to construct an enhanced feature matrix.

[0067] The attention weighting and dynamic threshold module, based on the enhanced feature matrix, uses an attention mechanism to assign weights to key behavioral features and sets a dynamic threshold adjustment mechanism to dynamically adjust the abnormal behavior judgment threshold according to the personnel density of the current scene.

[0068] The abnormal behavior recognition and reporting module is used to input adjusted feature data into the abnormal behavior classifier for recognition and judgment, output abnormal behavior recognition results, and generate abnormal event reports.

[0069] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the above-described edge AI-based video surveillance abnormal behavior recognition method.

[0070] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the above-described edge AI-based video surveillance abnormal behavior recognition method.

[0071] Compared with existing technologies, the advantages of this invention are as follows: By constructing a complete technical chain integrating target trajectory prediction, multi-scale feature fusion, adaptive attention mechanism, and intelligent classification and recognition, high-precision real-time recognition of abnormal behavior in video surveillance is achieved; the traditional static threshold judgment is transformed into a dynamic adaptive adjustment mechanism based on scene density, key areas are locked in advance by the target trajectory prediction model, behavioral semantic information of different granularities is captured by the multi-scale feature extraction network, and the attention mechanism is used to highlight the weight of key features. Finally, accurate abnormal event recognition is achieved through a deep separable convolution classifier. It fully considers the computing resource limitations of edge AI devices, reduces computational redundancy through prediction and localization, avoids environmental interference through dynamic thresholds, and improves recognition accuracy through feature fusion, forming an optimal balance between computational efficiency and recognition performance. It not only significantly improves the accuracy of abnormal behavior detection in complex monitoring scenarios and reduces the false alarm and false negative rates, but more importantly, it achieves millisecond-level real-time response capability, providing a technical solution with strong practical value for the field of smart security. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. 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. Wherein:

[0073] Figure 1 This is a flowchart of a video surveillance abnormal behavior recognition method based on edge AI. Detailed Implementation

[0074] 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.

[0075] 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.

[0076] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0077] As mentioned in the background section, existing video surveillance abnormal behavior recognition technologies still suffer from the following key problems: insufficient accuracy in target trajectory prediction, leading to tracking failure in occluded or fast-moving scenarios; simplistic feature fusion strategies that fail to fully exploit the correlation between spatiotemporal features; and anomaly detection relying on fixed thresholds and lacking dynamic adjustment mechanisms, thus affecting the system's generalization ability. To address these issues, this paper proposes an abnormal behavior recognition method and system based on multi-scale feature fusion and dynamic threshold adjustment using edge AI.

[0078] Figure 1 This is a flowchart illustrating a video surveillance abnormal behavior recognition method based on edge AI according to an embodiment of the present invention. Figure 1 As shown, a video surveillance abnormal behavior recognition method based on edge AI includes:

[0079] S1: Perform frame segmentation processing on the video surveillance data stream to obtain the video frame sequence. At the same time, establish a target trajectory prediction model, predict the location area of ​​the current frame based on the target location information, and generate target location prediction data.

[0080] Specifically, real-time video surveillance data streams are acquired from video surveillance equipment, and the video surveillance data streams are format converted and preprocessed to obtain preprocessed video surveillance data streams. The preprocessed video surveillance data streams are then segmented along the time axis according to a preset frame rate, decomposing the video surveillance data streams into discrete single-frame images to construct a video frame sequence.

[0081] It should be noted that the video surveillance data stream is received in real time through AI devices deployed at the edge; preprocessing converts the raw video surveillance data stream into a standardized digital video format, removing image noise and transmission interference; the video surveillance data stream includes continuous time-series image information; constructing a video frame sequence includes: numbering and marking the segmented single-frame images according to time order; each video frame in the video frame sequence has a unique time identifier.

[0082] Furthermore, target detection processing is performed on the video frames of the video frame sequence to identify personnel and object targets in the video frames. A bounding box localization algorithm is used to determine the position coordinates of each target in the current video frame, generating a target location information dataset. Based on the target location information dataset, a target tracking algorithm is used to perform identity association matching on the same target in adjacent video frames to construct target trajectory data, which includes target identity identifiers and historical location sequences.

[0083] Preferably, the target location information data includes center point coordinates, bounding box size, and confidence information; the target trajectory data is established by calculating feature similarity and constraining positional continuity to establish the trajectory association relationship of the target in consecutive video frames.

[0084] Furthermore, based on the target trajectory data, a target trajectory prediction model is established, in which the target trajectory prediction model adopts the Kalman filter algorithm; target position information of N consecutive frames is extracted from the target trajectory data to form a target historical position sequence input vector.

[0085] It should be noted that the target trajectory prediction model uses the target's historical position sequence as input features and learns the target's motion pattern and velocity change law through motion state estimation and trajectory trend analysis.

[0086] Preferably, the position information of the first N frames includes the center coordinates, speed and direction of motion of the target in the first N time steps.

[0087] Specifically, the target's historical location sequence input vector is input into the target trajectory prediction model. The model inferences and calculates to predict the target's location region in the current frame, generating a location prediction result. Based on the location prediction result, target location prediction data is generated, and the target location prediction data is bound with the corresponding target identity identifier to form a target location prediction dataset.

[0088] It should be noted that the target trajectory prediction model outputs the predicted location coordinates, prediction confidence, and range of location uncertainty; the location prediction results include the spatial location distribution probability.

[0089] For example, real-time frame segmentation is performed on the video stream, decomposing the continuous video stream into discrete single-frame images and adding a precise timestamp to each frame. Frame segmentation: The edge AI box receives the RTSP stream, uses FFmpeg to convert the video to H.264 format, and performs Gaussian filtering for noise reduction; it is segmented at 25fps to generate a frame sequence (e.g., frame_0001.jpg~frame_1000.jpg), with each frame appended with a millisecond-level timestamp (e.g., 20230714_10:00:00.042); trajectory prediction... Test: Detection: YOLOv5 identifies 3 pedestrians in frame 100 (boundary box coordinates such as (x1, y1, x2, y2) = (120, 80, 180, 200)); Tracking: DeepSORT associates the ID1 pedestrians in frames 100-105 and generates a trajectory sequence [(x = 120, y = 140), (125, 138), (130, 135)...]; Prediction: Kalman filtering uses the positions of the previous 5 frames as input and predicts the position of the 106th frame as (135, 132) ± 5px with a confidence level of 0.92.

[0090] S2: Perform multi-scale feature extraction on the video frame sequence to obtain the target feature vector, and fuse and match the target feature vector with the target position prediction data to construct an enhanced feature matrix.

[0091] Specifically, a convolutional kernel group is constructed, and the number of channels in the convolutional kernel group is configured to increase incrementally to form a multi-scale feature extraction network architecture. The convolutional kernel group includes small-sized convolutional kernels, medium-sized convolutional kernels, and large-sized convolutional kernels. The pixel values ​​of the video frame sequence are normalized to map the pixel value range to a standard numerical range, thereby generating standardized video frame sequence data.

[0092] It should be noted that the convolutional kernel group is used to extract visual feature information at different spatial scales; the multi-scale feature extraction network architecture can simultaneously capture fine-grained texture features and coarse-grained semantic features; and the video frame sequence is size-normalized to uniformly adjust the resolution of each video frame to the preset resolution specification.

[0093] Furthermore, standardized video frame sequence data is input into a multi-scale feature extraction network architecture, and features are extracted through convolutional kernel groups to obtain fine-grained feature maps, medium-grained feature maps, and coarse-grained feature maps.

[0094] Preferably, local detail features are extracted using small-sized convolutional kernels to obtain fine-grained feature maps; medium-scale structural features are extracted using medium-sized convolutional kernels to obtain medium-grained feature maps; and global semantic features are extracted using large-sized convolutional kernels to obtain coarse-grained feature maps.

[0095] Furthermore, channel-dimensional concatenation is performed on the fine-grained, medium-grained, and coarse-grained feature maps to generate a multi-scale fused feature map. Dimensionality reduction is applied to the multi-scale fused feature map using a dimensionality-reducing convolution kernel to compress the number of feature channels to a preset dimension, resulting in a compressed multi-scale feature map. Based on the prediction center coordinates and prediction region boundary information of the target location prediction data, the spatial region where the target is located is located on the compressed multi-scale feature map, and the RoI pooling algorithm is used to extract the feature sub-map corresponding to the target region.

[0096] It should be noted that the feature sub-image size is uniformly standardized; global average pooling is performed on the feature sub-image to convert the two-dimensional feature sub-image into a one-dimensional feature vector; a fully connected layer is used to perform a non-linear transformation on the one-dimensional feature vector to generate a high-dimensional target feature vector; the target feature vector contains the target's appearance features, shape features, and texture features.

[0097] Specifically, the prediction center coordinates, prediction region boundaries, and prediction confidence information of the target location prediction data are numerically encoded to generate a location encoding vector; the target feature vector and the location encoding vector are concatenated to generate a high-dimensional fusion feature vector, which serves as the enhanced feature vector.

[0098] It should be noted that the predicted center coordinates are converted into relative position codes by coordinate normalization, and the predicted region boundary is converted into scale codes by logarithmic transformation, generating a fixed-dimensional position code vector; feature interaction learning is performed on the fused feature vector through a multilayer perceptron; the multilayer perceptron contains multiple hidden layers; the number of neurons in the hidden layers is configured in decreasing order.

[0099] Preferably, the enhanced feature vector dimension is a preset feature dimension; based on the personnel and object targets in the video frame sequence, the corresponding enhanced feature vectors are obtained, and the enhanced feature vectors of all targets are arranged according to the target identity to construct an enhanced feature matrix; the number of rows in the enhanced feature matrix is ​​equal to the number of targets, and the number of columns is equal to the enhanced feature vector dimension.

[0100] For example, the multi-scale feature fusion input is as follows: the 106th frame is normalized to 640×640 resolution, and the pixel values ​​are normalized to [0, 1]; feature extraction: small convolutional kernel (3×3) extracts the details of the pedestrian's clothing texture in ID1 (fine-grained feature map contains stripe texture); medium convolutional kernel (5×5) captures the upper body posture (medium-grained feature map shows the arm raised); large convolutional kernel (7×7) identifies the overall direction of movement (coarse-grained feature map shows moving to the left); feature fusion: the predicted position of ID1 (135, 132) is encoded into a normalized vector [0.21, 0.21]; RoI pooling extracts the feature sub-map of the ID1 region (56×56 pixels), and global pooling is used to obtain a 512-dimensional feature vector; visual features and position encoding are concatenated to generate a 518-dimensional enhanced feature vector (e.g., [0.2, 0.7, ..., 0.21, 0.21]).

[0101] S3: Based on the enhanced feature matrix, an attention mechanism is used to assign weights to key behavioral features and a dynamic threshold adjustment mechanism is set to dynamically adjust the threshold for judging abnormal behavior according to the personnel density of the current scene.

[0102] Specifically, the enhanced feature matrix is ​​transformed through a linear transformation layer to generate a query matrix Q, a key matrix K, and a value matrix V, respectively; the matrix product of the query matrix Q and the transpose of the key matrix K is calculated to obtain the attention score matrix S; the elements of the attention score matrix S are divided by a scaling factor to obtain the scaled attention score matrix; the rows of the scaled attention score matrix are subjected to Softmax normalization to generate the attention weight matrix A.

[0103] It should be noted that the dimensions of the query matrix Q, key matrix K, and value matrix V are all the same: number of targets × feature dimension. The enhanced feature matrix is ​​linearly transformed by the weight matrices WQ, WK, and WV. The dimensions of the weight matrices WQ, WK, and WV are all feature dimension × feature dimension. The dimension of the attention score matrix S is the number of targets × number of targets. The scaling factor is equal to the square root of the feature dimension. The dimension of the attention weight matrix A is the number of targets × number of targets. Each element in the attention weight matrix A represents the attention weight between targets. The sum of the elements in each row of the attention weight matrix A is equal to the unit value.

[0104] Furthermore, the attention weight matrix A and the value matrix V are multiplied to obtain the weighted feature matrix W; the weighted feature matrix W and the enhancement feature matrix are then joined by a residual connection to generate the residual feature matrix R.

[0105] Preferably, the dimension of the weighted feature matrix W is the number of targets × the feature dimension; each row of the weighted feature matrix W represents the target feature vector after being weighted by the attention mechanism; the residual feature matrix R is equal to the element-wise addition of the weighted feature matrix W and the enhanced feature matrix; the residual feature matrix R is subjected to layer normalization to obtain the normalized feature matrix, and the mean of each row of the normalized feature matrix is ​​zero and the variance is a unit value.

[0106] Furthermore, based on the target location prediction data, the number of personnel targets and the area of ​​the monitored area in the current video frame are calculated to obtain the personnel count and personnel density values. A dynamic mapping function between personnel density and abnormal behavior judgment threshold is established. The personnel density values ​​are divided into level intervals, and corresponding threshold adjustment coefficients are set for the density levels. A baseline abnormal behavior judgment threshold is set, and the personnel density values ​​are input into the dynamic mapping function to obtain the threshold adjustment coefficients corresponding to the current scene. The abnormal behavior judgment threshold is generated through numerical calculation. The residual feature matrix R and the abnormal behavior judgment threshold are correlated, and the feature weights of the residual feature matrix are adjusted a second time to generate feature data adjusted by the dynamic threshold, which is used as the input of the abnormal behavior classifier.

[0107] It should be noted that the personnel density value is calculated by dividing the personnel count by the area of ​​the monitored area; the personnel density value represents the number of people per unit area.

[0108] Preferably, density level thresholds are set to divide the area into intervals; when the personnel density value is less than the first threshold, it is classified as a first-level density level, and the anomaly detection threshold is set to the first-level threshold; when the personnel density value is between the first and second thresholds, it is classified as a second-level density level, and the anomaly detection threshold is set to the second-level threshold; when the personnel density value is between the second and third thresholds, it is classified as a third-level density level, and the anomaly detection threshold is set to the third-level threshold; when the personnel density value is greater than the third threshold, it is classified as a fourth-level density level, and the anomaly detection threshold is set to the fourth-level threshold.

[0109] It should be noted that the first threshold is set based on the accuracy of abnormal behavior recognition in low-density scenarios; the second threshold is set based on the degree of target occlusion interference in medium-density scenarios; the third threshold is set based on the complexity of group behavior in high-density scenarios; and the fourth threshold is set based on the safety risk warning requirements in extremely high-density scenarios, where the population density reaches a crowded state, and any abnormal behavior may trigger a chain reaction. The abnormal behavior judgment threshold is set to the lowest value to maximize the coverage of abnormal behavior detection.

[0110] For example, a dynamic threshold adjustment scenario: a sudden crowd gathering occurs on a subway platform (50 people are detected in a 20㎡ monitoring area); density calculation: people density = 50 people / 20㎡ = 2.5 people / ㎡ → triggering level 4 density (threshold > 2.0 people / ㎡); threshold adjustment: baseline threshold 0.8 → level 4 coefficient 0.5 → dynamic threshold = 0.8 × 0.5 = 0.4; after attention mechanism weighting, the ID47 feature vector (waving action) score is 0.45 > 0.4, and it is retained as a potential anomaly.

[0111] S4: Input the adjusted feature data into the abnormal behavior classifier for identification and judgment, output the abnormal behavior identification results, and generate an abnormal event report.

[0112] Specifically, an abnormal behavior classifier is constructed, which adopts a depthwise separable convolutional structure, including a feature encoding layer, a behavior pattern recognition layer, and an anomaly detection layer.

[0113] It should be noted that the feature encoding layer receives feature data that has been dynamically thresholded and performs dimensionality transformation and feature compression on the feature data; the behavior pattern recognition layer performs time series modeling on the feature data through temporal convolution kernels to capture the target's motion trajectory and behavior pattern features; and the anomaly detection layer classifies and judges the behavior pattern features based on preset abnormal behavior category labels.

[0114] Furthermore, the feature data adjusted by dynamic thresholds is grouped and organized according to the target identity to construct a target feature dataset. The target feature dataset is then temporally segmented to generate fixed-length feature sequence fragments. These feature sequence fragments are input into the feature encoding layer, where a one-dimensional convolution kernel is used to perform convolution operations on the feature sequence fragments to extract local feature patterns. The encoded feature vectors are then transmitted to the behavior pattern recognition layer, where a bidirectional long short-term memory network is used to perform temporal modeling on the encoded feature vectors to capture the target's behavioral continuity features and motion trend features.

[0115] It should be noted that the target feature dataset is batch normalized to uniformly map the numerical range of the feature data to a standard interval; the feature sequence fragments serve as the input units of the abnormal behavior classifier; the feature encoding layer outputs encoded feature vectors, which include the target's spatial location information and motion state information.

[0116] Furthermore, the behavior feature representation vector is output through the behavior pattern recognition layer, and the behavior feature representation vector is compared with the abnormal behavior judgment threshold. The behavior feature representation vector after threshold filtering is input into the anomaly judgment layer, which includes classification heads for several abnormal behavior categories.

[0117] Preferably, if the value of the behavior feature representation vector exceeds the abnormal behavior judgment threshold, it is marked as a potential abnormal behavior; the behavior feature representation vector is filtered by a threshold masking mechanism to retain feature components that exceed the threshold and suppress feature components that are below the threshold.

[0118] It should be noted that each category head corresponds to a specific type of abnormal behavior, including fighting behavior, falling behavior, gathering behavior, and escape behavior.

[0119] Specifically, the behavior feature representation vector is linearly transformed by a fully connected layer to generate an abnormal behavior probability distribution vector; the abnormal behavior probability distribution vector is then normalized using Softmax to obtain the confidence score of each abnormal behavior category, and the abnormal behavior category with the highest confidence score is selected as the recognition result.

[0120] Preferably, the identification results include: establishing a multi-dimensional judgment mechanism based on the correlation between the first confidence score and the density level threshold; when the personnel density value is at the first-level density and the first confidence score exceeds the first-level threshold, simultaneously calculating the entropy value of the abnormal behavior probability distribution vector; if the entropy value is lower than the preset entropy value threshold, then confirming the abnormal behavior identification result as a high-confidence abnormal event; when the personnel density value is at the second-level density and the first confidence score is between the second-level threshold and the first-level threshold, then extracting the temporal change gradient of the behavior feature representation vector; if the temporal change gradient exceeds the preset gradient threshold, then marking the abnormal behavior identification result as a gradual abnormal event; when the difference between the first confidence score and the second confidence score is less than the preset confidence competition threshold, then triggering the abnormal behavior category competition judgment.

[0121] Specifically, the judgment of abnormal behavior categories includes: if the population density value is at level 3 and the first confidence score corresponds to gathering behavior and the second confidence score corresponds to escape behavior, it is judged as an abnormal crowd evacuation event; if the population density value is at level 4 and the first confidence score corresponds to fighting behavior and the second confidence score corresponds to gathering behavior, it is judged as an abnormal group conflict event; when the confidence competition threshold condition is met and the population density value is at level 1, the abnormal behavior category corresponding to the first confidence score is selected as the identification result.

[0122] Furthermore, a mapping relationship is established between density levels and interaction response levels. When the personnel density value is at level 1 and an abnormal behavior event is detected, a level 1 interaction response is triggered, immediately pushing the abnormal behavior identification result to the monitoring terminal and activating the audible and visual alarm device. When the personnel density value is at level 2 and the first confidence score exceeds the level 2 threshold, a level 2 interaction response is triggered, sending the abnormal behavior identification result to the management platform and recording the abnormal event log. When the personnel density value is at level 3 and the first confidence score exceeds the level 3 threshold, a level 3 interaction response is triggered, recording the abnormal behavior identification result only in the system background without real-time alarm. When the personnel density value is at level 4, due to the excessively high personnel density, the interaction response level of the abnormal behavior identification result is automatically upgraded by one level.

[0123] Furthermore, if an abnormal behavior event is detected, the spatiotemporal location information of the abnormal target is extracted from the residual feature matrix R, and the spatiotemporal location information is associated and bound with the abnormal behavior identification result to generate an abnormal behavior event record. Based on the abnormal behavior event record, an abnormal event report is constructed, and the abnormal event report is sent to the monitoring center or mobile terminal through the communication module of the edge AI device to complete the rapid detection and alarm process of abnormal behavior.

[0124] It should be noted that the spatiotemporal location information includes the target's pixel coordinates, bounding box coordinates, and timestamp information in the current video frame; the abnormal event report includes the abnormal behavior category, occurrence time, location coordinates, target identification, confidence score, and duration.

[0125] For example, the abnormal behavior classification input is a 10-frame feature sequence of ID47 (wave frequency 2Hz); the classification process is a bidirectional LSTM to identify periodic wave patterns; the output probabilities of the fully connected layer are: fighting: 0.62, falling: 0.05, gathering: 0.28, escaping: 0.05; the entropy value is calculated to be 0.85 (below the threshold of 1.2), confirming it as a high-confidence fighting event; the response mechanism is as follows: under the fourth density, it should be a third-level response, but it is automatically upgraded to the second level: push alarm information to the security terminal: fighting behavior, ID47, location (320, 480), confidence level 62%; and simultaneously save the event log and video clips (frames 106-116).

[0126] In summary, this invention achieves high-precision real-time identification of abnormal behavior in video surveillance by constructing a complete technical chain integrating target trajectory prediction, multi-scale feature fusion, adaptive attention mechanism, and intelligent classification and recognition. It transforms traditional static threshold judgment into a dynamic adaptive adjustment mechanism based on scene density. Key areas are pre-identified through a target trajectory prediction model, and behavioral semantic information of different granularities is captured by a multi-scale feature extraction network. An attention mechanism is then used to highlight the weights of key features, and finally, a deep separable convolutional classifier achieves accurate abnormal event identification. This invention fully considers the computational resource limitations of edge AI devices, reduces computational redundancy through predictive localization, avoids environmental interference through dynamic thresholds, and improves recognition accuracy through feature fusion, achieving an optimal balance between computational efficiency and recognition performance. This not only significantly improves the accuracy of abnormal behavior detection in complex monitoring scenarios and reduces false alarms and missed alarms, but more importantly, it achieves millisecond-level real-time response capabilities, providing a highly practical technical solution for the field of smart security.

[0127] Furthermore, this embodiment also provides a video surveillance abnormal behavior recognition system based on edge AI, including: a frame processing and trajectory prediction module, used to perform frame segmentation processing on the video surveillance data stream to obtain a video frame sequence, and simultaneously establish a target trajectory prediction model to predict the location region of the current frame based on the target location information, generating target location prediction data; a feature extraction and fusion module, used to perform multi-scale feature extraction on the video frame sequence to obtain a target feature vector, and fuse and match the target feature vector with the target location prediction data to construct an enhanced feature matrix; an attention weighting and dynamic threshold module, based on the enhanced feature matrix, using an attention mechanism to assign weights to key behavioral features, setting a dynamic threshold adjustment mechanism, and dynamically adjusting the abnormal behavior judgment threshold according to the personnel density of the current scene; and an abnormal behavior recognition and reporting module, used to input the adjusted feature data into an abnormal behavior classifier for recognition and judgment, output the abnormal behavior recognition result, and generate an abnormal event report.

[0128] This embodiment also provides a computer device applicable to the video surveillance abnormal behavior recognition method based on edge AI, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the video surveillance abnormal behavior recognition method based on edge AI as proposed in the above embodiment.

[0129] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0130] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the video surveillance abnormal behavior recognition method based on edge AI as proposed in the above embodiments.

[0131] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0132] 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 method for identifying abnormal behavior in video surveillance based on edge AI, characterized in that: include, The video surveillance data stream is segmented to obtain a video frame sequence. At the same time, a target trajectory prediction model is established to predict the location region of the current frame based on the target location information, and target location prediction data is generated. Multi-scale feature extraction is performed on the video frame sequence to obtain the target feature vector, and the target feature vector is fused and matched with the target position prediction data to construct an enhanced feature matrix; Based on the enhanced feature matrix, an attention mechanism is used to assign weights to key behavioral features, and a dynamic threshold adjustment mechanism is set to dynamically adjust the abnormal behavior judgment threshold according to the personnel density of the current scene. The adjusted feature data is input into the abnormal behavior classifier for identification and judgment, and the abnormal behavior identification results are output to generate an abnormal event report.

2. The video surveillance abnormal behavior recognition method based on edge AI as described in claim 1, characterized in that: The adjusted feature data is input into the abnormal behavior classifier for identification and judgment, and the abnormal behavior identification results are output, generating an abnormal event report, including: Construct an abnormal behavior classifier, wherein the abnormal behavior classifier adopts a depthwise separable convolutional structure, including a feature encoding layer, a behavior pattern recognition layer, and an anomaly determination layer; The feature data adjusted by dynamic thresholds are grouped and organized according to the target identity to construct a target feature dataset. The target feature dataset is then segmented temporally to generate fixed-length feature sequence fragments. The feature sequence fragment is input into the feature encoding layer, and a one-dimensional convolution kernel is used to perform a convolution operation on the feature sequence fragment to extract local feature patterns. The encoded feature vector is transmitted to the behavior pattern recognition layer, and a bidirectional long short-term memory network is used to perform temporal modeling on the encoded feature vector to capture the target's behavioral continuity features and motion trend features. The behavior pattern recognition layer outputs a behavior feature representation vector, and the behavior feature representation vector is compared with the abnormal behavior determination threshold. The behavior feature representation vector that has been filtered by the threshold is input into the anomaly determination layer, wherein the anomaly determination layer includes a classification head for several abnormal behavior categories; A fully connected layer is used to perform a linear transformation on the behavior feature representation vector to generate an abnormal behavior probability distribution vector. The probability distribution vector of the abnormal behavior is subjected to Softmax normalization to obtain the confidence score of each abnormal behavior category, and the abnormal behavior category with the first confidence score is selected as the recognition result. If an abnormal behavior event is detected, the spatiotemporal location information of the abnormal target is extracted from the residual feature matrix R, and the spatiotemporal location information is associated and bound with the abnormal behavior identification result to generate an abnormal behavior event record. Based on the abnormal behavior event records, an abnormal event report is constructed, and the abnormal event report is sent to the monitoring center or mobile terminal through the communication module of the edge AI device to complete the rapid detection and alarm process of abnormal behavior.

3. The video surveillance abnormal behavior recognition method based on edge AI as described in claim 2, characterized in that: The identification results include: A multi-dimensional judgment mechanism is established based on the correlation between the first confidence score and the density level threshold. When the personnel density value is at the first-level density level and the first confidence score exceeds the first-level threshold, the entropy value of the abnormal behavior probability distribution vector is calculated at the same time; if the entropy value is lower than the preset entropy value threshold, the abnormal behavior identification result is confirmed as a high-confidence abnormal event. When the personnel density value is at the second-level density level and the first confidence score is between the second-level threshold and the first-level threshold, the temporal change gradient of the behavior feature representation vector is extracted; if the temporal change gradient exceeds the preset gradient threshold, the abnormal behavior identification result is marked as a gradual abnormal event. When the difference between the first confidence score and the second confidence score is less than the preset confidence competition threshold, the abnormal behavior category competition judgment is triggered.

4. The video surveillance abnormal behavior recognition method based on edge AI as described in claim 3, characterized in that: The judgment of the competition of abnormal behavior categories includes: If the population density value is at level three and the first confidence score corresponds to gathering behavior and the second confidence score corresponds to escape behavior, then it is determined to be an abnormal crowd evacuation event. If the population density value is at level four and the first confidence score corresponds to fighting behavior and the second confidence score corresponds to gathering behavior, then it is determined to be an abnormal group conflict event. When the confidence threshold condition is met and the personnel density value is at the first density level, the abnormal behavior category corresponding to the first confidence score is selected as the identification result.

5. The video surveillance abnormal behavior recognition method based on edge AI as described in claim 3, characterized in that: Based on the enhanced feature matrix, an attention mechanism is used to assign weights to key behavioral features, and a dynamic threshold adjustment mechanism is set to dynamically adjust the abnormal behavior judgment threshold according to the personnel density of the current scene, including: The enhanced feature matrix is ​​transformed into a query matrix Q, a key matrix K, and a value matrix V through a linear transformation layer. Calculate the matrix product of the query matrix Q and the transpose of the key matrix K to obtain the attention score matrix S; Divide each element of the attention score matrix S by a scaling factor to obtain the scaled attention score matrix; The rows of the scaled attention score matrix are subjected to Softmax normalization to generate the attention weight matrix A; Perform a matrix multiplication operation between the attention weight matrix A and the value matrix V to obtain the weighted feature matrix W; Perform a residual concatenation operation between the weighted feature matrix W and the enhanced feature matrix to generate a residual feature matrix R; Based on the target location prediction data, calculate the number of people in the current video frame and the area of ​​the monitored area to obtain the people count and people density values. Establish a dynamic mapping function between personnel density and abnormal behavior judgment threshold, divide the level range according to the personnel density value, and set a corresponding threshold adjustment coefficient for the density level; Set a baseline abnormal behavior judgment threshold, input the personnel density value into the dynamic mapping function, obtain the threshold adjustment coefficient corresponding to the current scene, and generate the abnormal behavior judgment threshold through numerical calculation; The residual feature matrix R and the abnormal behavior judgment threshold are correlated, and the feature weights of the residual feature matrix R are adjusted a second time to generate feature data that has been dynamically thresholded, which is then used as the input to the abnormal behavior classifier.

6. The video surveillance abnormal behavior recognition method based on edge AI as described in claim 5, characterized in that: The method for establishing the enhanced feature matrix is ​​as follows: Convolutional kernel groups are constructed, and the number of channels in the convolutional kernel groups is configured to increase incrementally to form a multi-scale feature extraction network architecture, wherein the convolutional kernel groups include small-sized convolutional kernels, medium-sized convolutional kernels, and large-sized convolutional kernels; The pixel values ​​of the video frame sequence are normalized to map the range of pixel values ​​to a standard numerical range, thereby generating standardized video frame sequence data. The standardized video frame sequence data is input into the multi-scale feature extraction network architecture, and features are extracted through the convolutional kernel group to obtain fine-grained feature maps, medium-grained feature maps, and coarse-grained feature maps; The fine-grained feature map, medium-grained feature map, and coarse-grained feature map are concatenated along the channel dimension to generate a multi-scale fused feature map. The multi-scale fused feature map is reduced in dimensionality by using a dimensionality-reducing convolution kernel to compress the number of feature channels to a preset dimension, thereby obtaining a compressed multi-scale feature map. Based on the predicted center coordinates and predicted region boundary information of the target location prediction data, the spatial region where the target is located is located on the compressed multi-scale feature map, and the RoI pooling algorithm is used to extract the feature sub-map corresponding to the target region. The prediction center coordinates, prediction area boundaries, and prediction confidence information of the target location prediction data are numerically encoded to generate a location encoding vector. The target feature vector and the position encoding vector are concatenated to generate a high-dimensional fusion feature vector, which is used as the enhanced feature vector. Based on the video frame sequence of personnel and object targets, the corresponding enhanced feature vectors are obtained, and the enhanced feature vectors of all targets are arranged according to the target identity to construct an enhanced feature matrix.

7. The video surveillance abnormal behavior recognition method based on edge AI as described in claim 6, characterized in that: The video surveillance data stream is segmented to obtain a video frame sequence. Simultaneously, a target trajectory prediction model is established to predict the location region of the current frame based on target location information, generating target location prediction data, including: The system acquires real-time video surveillance data streams from video surveillance equipment, performs format conversion and preprocessing on the video surveillance data streams, and obtains preprocessed video surveillance data streams. The preprocessed video surveillance data stream is divided into time-axis segments according to a preset frame rate, and the video surveillance data stream is decomposed into discrete single-frame images to construct a video frame sequence. The video frames of the video frame sequence are subjected to target detection processing to identify personnel targets and object targets in the video frames, and the bounding box localization algorithm is used to determine the position coordinates of each target in the current video frame to generate a target position information dataset. Based on the target location information dataset, a target tracking algorithm is used to perform identity association matching on the same target in adjacent video frames to construct target trajectory data, wherein the target trajectory data includes target identity identifiers and historical location sequences; Based on the target trajectory data, a target trajectory prediction model is established, wherein the target trajectory prediction model adopts the Kalman filter algorithm; Extract target location information from N consecutive frames of the target trajectory data to form a target historical location sequence input vector; The target historical location sequence input vector is input into the target trajectory prediction model to calculate and predict the target's location region in the current frame, and generate a location prediction result. Based on the location prediction results, target location prediction data is generated, and the target location prediction data is bound with the corresponding target identity identifier to form a target location prediction dataset.

8. A video surveillance abnormal behavior recognition system based on edge AI, based on the video surveillance abnormal behavior recognition method based on edge AI according to any one of claims 1 to 7, characterized in that: include, The frame processing and trajectory prediction module is used to perform frame segmentation processing on the video surveillance data stream to obtain the video frame sequence. At the same time, it establishes a target trajectory prediction model, predicts the location region of the current frame based on the target location information, and generates target location prediction data. The feature extraction and fusion module is used to extract multi-scale features from the video frame sequence to obtain target feature vectors, and to fuse and match the target feature vectors with the target position prediction data to construct an enhanced feature matrix. The attention weighting and dynamic threshold module, based on the enhanced feature matrix, uses an attention mechanism to assign weights to key behavioral features and sets a dynamic threshold adjustment mechanism to dynamically adjust the abnormal behavior judgment threshold according to the personnel density of the current scene. The abnormal behavior recognition and reporting module is used to input adjusted feature data into the abnormal behavior classifier for recognition and judgment, output abnormal behavior recognition results, and generate abnormal event reports.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the video surveillance abnormal behavior recognition method based on edge AI as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the video surveillance abnormal behavior recognition method based on edge AI as described in any one of claims 1 to 7.

Images