An analysis method and system based on AI video recognition behavior monitoring

Abstract

This invention discloses an analysis method and system based on AI video recognition behavior monitoring, belonging to the field of video recognition technology. The method includes: acquiring video data, calculating the overall pixel gradient change of adjacent frames, marking preliminary candidate frames, calculating the edge saliency intensity of each frame, analyzing the set of salient region points, performing clustering, calculating the geometric center point of each cluster as an anchor value, and forming an anchor point set using the center points of each cluster. The method uses the density-based clustering algorithm DBSCAN to achieve centralized extraction of salient points, ensuring the logical uniqueness and local integrity of the extracted results. It uses a shallow convolutional network to extract local feature maps, effectively capturing target texture and edge information, and optimizes the bounding box using a rotation IoU loss function, improving the fitting accuracy of salient region boundaries.

Description

Technical Field

[0001] This invention relates to the field of video recognition technology, and in particular to an analysis method and system based on AI-based video recognition behavior monitoring. Background Technology

[0002] Traditional video surveillance systems typically rely on fixed rules and simple algorithms to process video data. In recent years, deep learning and artificial intelligence methods have been widely applied to video analysis, especially in the field of behavior recognition. By introducing neural network models, more robust target detection and classification have been achieved. AI-based video recognition technology has gradually expanded from single-frame image detection to time-series analysis of continuous frames. This technology can combine spatial characteristics and temporal dynamic information to accurately identify and analyze complex behaviors in real time.

[0003] However, with the diversification of application scenarios, the size, position, and shape of target behavior in videos may change dynamically. Traditional video target detection methods mostly extract target features based on fixed bounding boxes or preset regions. When the target changes angle or motion shape in the video, the fixed bounding box cannot adapt effectively, resulting in a decrease in recognition accuracy. Secondly, although neural network models can improve the accuracy of behavior classification, when presenting dynamic abnormal behavior, existing classification methods do not fully combine the statistical characteristics of salient regions and cannot effectively detect the subtle fluctuations of specific abnormal targets. In addition, the combination of shallow and deep features is rarely used. Most systems only extract local features based on deep features, while ignoring the important role of low-level spatial characteristics in the accurate localization of behavior regions, which easily leads to insufficient model robustness when analyzing multi-target behavior in complex scenes. Summary of the Invention

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

[0005] Therefore, this invention provides an analysis method and system based on AI video recognition behavior monitoring. Traditional video target detection methods often rely on fixed bounding boxes or preset regions to extract target features. When the target's angle or motion changes in the video, the fixed bounding box cannot effectively adapt, leading to decreased recognition accuracy. Secondly, while neural network models can improve behavior classification accuracy, existing classification methods do not fully incorporate the statistical characteristics of salient regions when presenting dynamic abnormal behavior, failing to effectively detect subtle fluctuations in specific abnormal targets. Furthermore, the combination of shallow and deep features is rarely used; most systems only extract local features based on deep features, neglecting the crucial role of underlying spatial characteristics in accurate behavior region localization. This easily leads to insufficient model robustness when analyzing multi-target behavior in complex scenes.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an analysis method based on AI video recognition behavior monitoring, comprising:

[0008] Collect video data to calculate the overall pixel gradient change of adjacent frames, mark preliminary candidate frames to calculate the edge saliency intensity of each frame, analyze the set of salient region points, perform clustering to calculate the geometric center point of each cluster as the anchor point value, and form an anchor point set through the center point of each cluster.

[0009] For each anchor point coordinate, a local segmentation window is defined to form a set of local windows. A shallow convolutional network is used to generate feature maps for local feature extraction. Figure 2 The binary mask is calculated by quantization, the vertex coordinates of the rectangular bounding box are defined, the rectangular bounding box is determined, and the rotation IoU loss function is used as the evaluation index of the bounding box fitting quality to dynamically optimize the rectangular bounding box.

[0010] Calculate the attention weight map of the salient region, perform enhancement calculation on the channel weights, generate a set of channel-level enhanced feature maps, predict the classification label corresponding to the highest probability through a pre-trained feature classification model, perform gradient calculation on the feature map according to the loss function, reallocate the channel-level attention weights, and update the channel weights.

[0011] Extract neighborhood features of the target behavior region, calculate the variance of the channel features to generate abnormal feature responses, and analyze abnormal targets.

[0012] As a preferred embodiment of the AI-based video recognition behavior monitoring analysis method of the present invention, wherein: the geometric center point of each cluster is calculated as an anchor point value, and an anchor point set is formed by the center points of each cluster, including:

[0013] Video data is acquired, and consecutive frames of video image data are combined into an image sequence. The overall pixel gradient change between adjacent frames is calculated using the Frobenius norm.

[0014] Based on historical experience, a preliminary screening threshold is determined, and frames with a global pixel change rate greater than or equal to the preliminary screening threshold are marked as preliminary candidate frames.

[0015] For the initial candidate frame set, Gaussian blur is applied for denoising, and the Sobel operator is used to calculate the edge saliency intensity and edge saliency matrix of each frame;

[0016] Based on the sum of the mean and twice the standard deviation of the edge saliency intensity in historical experience as the region threshold, pixels with edge saliency intensity greater than or equal to the region threshold are regarded as the set of salient region points.

[0017] The density-based spatial clustering algorithm DBSCAN clusters salient region points, calculates the Euclidean distance between salient region points, determines the average neighborhood distance as the neighborhood radius for each salient region point, calculates the minimum density number of points based on the total number of salient region points, and then calculates the geometric center point of each cluster as the anchor point value. An anchor point set is formed by the center points of each cluster.

[0018] As a preferred embodiment of the AI-based video recognition behavior monitoring analysis method of the present invention, the steps of defining the vertex coordinates of a rectangular bounding box, determining the rectangular bounding box, using the rotation IoU loss function as an evaluation index for the bounding box fitting quality, and dynamically optimizing the rectangular bounding box include:

[0019] Based on the set of anchor points, a local segmentation window is defined for the coordinates of each anchor point to obtain local image patches centered on the anchor points, which are then combined into a set of local windows.

[0020] For each local image patch, a shallow convolutional network is used to generate a feature map, which is then normalized.

[0021] Based on the local image patches corresponding to the normalized shallow features, perform local feature extraction. Figure 2 The binary mask is calculated by quantization, and the maximum and minimum pixel coordinates are determined. The vertex coordinates of the rectangular bounding box are defined to determine the rectangular bounding box.

[0022] The rotation IoU loss function is used as an evaluation index for the quality of bounding box fitting. The rectangular bounding box is dynamically optimized, and the key parameters of the bounding box are gradually adjusted using the gradient descent method.

[0023] As a preferred embodiment of the AI-based video recognition behavior monitoring analysis method of the present invention, wherein: the calculation of the attention weight map of the salient region, the enhancement calculation of the channel weights, and the generation of a set of channel-level enhanced feature maps, including,

[0024] For dynamically optimized rectangular bounding boxes, regional images are cut from the original video frames, attention weight maps of salient regions are calculated, and target features of salient regions are optimized through attention weight allocation mechanisms to generate pixel-level salient attention feature maps.

[0025] For each salient attention feature map, global average pooling is performed on each channel to generate statistical features of the channel position. The sigmoid activation function is used to generate attention weights for each channel. Enhancement calculations are performed on the channel weights to generate a set of channel-level enhanced feature maps.

[0026] As a preferred embodiment of the AI-based video recognition behavior monitoring analysis method of the present invention, the steps of generating a set of channel-level enhanced feature maps, predicting the classification label corresponding to the highest probability using a pre-trained feature classification model, calculating the gradient of the feature maps according to the loss function, reallocating channel-level attention weights, and updating the channel weights include:

[0027] By using a pre-trained feature classification model, channel-level enhanced feature maps are input into the classifier ANN artificial neural network to obtain the classification label corresponding to the highest predicted probability. The supervision signal loss is calculated based on the true label and the predicted probability, and the loss function is the cross-entropy loss. The gradient of the feature map is calculated according to the loss function.

[0028] Based on the gradient update results, channel-level attention weights are reallocated and channel weights are updated. Based on the updated channel weight information, the enhanced feature map is recalculated to generate the final optimized feature set.

[0029] As a preferred embodiment of the AI-based video recognition behavior monitoring analysis method of the present invention, the step of extracting neighborhood features of the target behavior region and calculating the variance of channel features to generate anomaly feature responses includes:

[0030] From the final optimized feature set and the dynamically optimized rectangular bounding box, the neighborhood features of the target behavior region are extracted, and the variance of the channel features is calculated to generate the abnormal feature response.

[0031] As a preferred embodiment of the AI-based video recognition behavior monitoring analysis method of the present invention, wherein: the analyzed abnormal targets include,

[0032] The anomaly response threshold is determined by the sum of the mean and twice the standard deviation of historical anomaly response features. If the anomaly response feature is greater than or equal to the anomaly response threshold, it is judged as an anomaly target.

[0033] Secondly, the present invention provides an analysis system based on AI video recognition and behavior monitoring, comprising,

[0034] The video frame preprocessing module acquires video data, calculates the overall pixel gradient change between adjacent frames, and marks preliminary candidate frames with significant changes.

[0035] The salient region extraction module calculates the edge saliency intensity of each frame, analyzes the salient region point set, and extracts the geometric center of the salient point cluster through clustering to generate an anchor point set;

[0036] The target region boundary optimization module defines a local segmentation window based on anchor points and generates a rectangular bounding box. It then dynamically optimizes the spatial fit of the bounding box using a rotation IoU loss function.

[0037] The salient feature optimization module calculates pixel attention maps of salient regions and enhances channel weights to generate channel-level salient enhanced feature maps.

[0038] The supervised optimization module predicts classification labels using a pre-trained classification model, calculates the gradient of the supervision signal, and reallocates channel weights to optimize the feature map.

[0039] The anomaly detection module extracts neighborhood features of the target behavior region, calculates the variance of the channel features to generate anomaly response, and completes the analysis of the anomaly target.

[0040] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the analysis method based on AI video recognition behavior monitoring as described in the first aspect of the present invention.

[0041] Fourthly, the present invention provides 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 analysis method based on AI video recognition behavior monitoring as described in the first aspect of the present invention.

[0042] The beneficial effects of this invention are as follows: It achieves centralized extraction of salient points through the density-based clustering algorithm DBSCAN, ensuring the extraction results possess logical uniqueness and local integrity of salient targets. By using a shallow convolutional network to extract local feature maps, it effectively captures target texture and edge information. The bounding box optimization achieved through the rotation IoU loss function improves the fitting accuracy of salient region boundaries. The generation of a channel-level enhanced feature map set realizes multi-scale global and local fusion of salient region features, improving the robustness and adaptability of content features. Through gradient calculation, it dynamically adjusts the salient target feature map, accurately reshaping the characteristic response of the target region, avoiding the limitations of directly setting weights. Finally, the feature set forms highly expressive feature data that combines category characteristics with subtle information within salient regions. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating the analysis method based on AI video recognition behavior monitoring in Example 1.

[0045] Figure 2This is a schematic diagram of the analysis system based on AI video recognition behavior monitoring in Example 1. Detailed Implementation

[0046] 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.

[0047] 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.

[0048] 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.

[0049] Example 1, referring to Figures 1 to 2 This is the first embodiment of the present invention, which provides an analysis method based on AI video recognition behavior monitoring, including the following steps:

[0050] S1. Collect video data, calculate the overall pixel gradient change of adjacent frames, mark preliminary candidate frames, calculate the edge saliency intensity of each frame, analyze the set of salient region points, perform clustering calculation, use the geometric center point of each cluster as the anchor point value, and form an anchor point set through the center point of each cluster.

[0051] Preferably, the geometric center point of each cluster is used as the anchor point value for clustering calculation, and an anchor point set is formed by the center points of each cluster, including:

[0052] Video data is acquired, and consecutive video image data are combined into an image sequence. The overall pixel gradient change between adjacent frames is calculated using the Frobenius norm, and expressed as follows:

[0053]

[0054]

[0055] in, This represents the change in value of each pixel in time frame t. and These represent the pixel matrix data for time frames t+1 and t, respectively. Indicates the global pixel change rate. It represents the Frobenius norm of the matrix and measures the pixel gradient intensity throughout the frame;

[0056] Based on historical experience, a preliminary screening threshold is determined, and frames with a global pixel change rate greater than or equal to the preliminary screening threshold are marked as preliminary candidate frames.

[0057] For the initial candidate frame set, Gaussian blur is applied for denoising, and the Sobel operator is used to calculate the edge saliency intensity of each frame. The edge saliency matrix is ​​expressed as:

[0058]

[0059] in, This represents the pixel matrix data of time frame t after denoising. These represent the edge saliency intensity of pixels, forming an edge saliency matrix;

[0060] Based on the sum of the mean and twice the standard deviation of the edge saliency intensity in historical experience as the region threshold, pixels with edge saliency intensity greater than or equal to the region threshold are regarded as the set of salient region points.

[0061] The density-based spatial clustering algorithm DBSCAN clusters salient region points, calculates the Euclidean distance between salient region points, determines the average neighborhood distance as the neighborhood radius for each salient region point, and calculates the minimum density of points based on the total number of salient region points. Then, it calculates the geometric center point of each cluster as the anchor point value, and forms an anchor point set using the center points of each cluster, represented as follows:

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] in, Let Euclidean distances be the distances between points i and j in the salient regions, and let them form a distance matrix. and Let i be the coordinates of the i-th salient region point. and Let j be the coordinates of the j-th salient region point. This represents the local neighborhood radius of the salient region point i. This represents the set of distances from point i to all other points, extracted from all columns of the distance matrix in row i. This represents a sorting operation, where the k-th smallest distance value is taken from the set of distances from point i to all other points, where k represents the k-th nearest neighbor of point i. The radius represents the global neighborhood, the distance radius used to define the cluster, and N represents the total number of salient points. The value represents the estimated number of points covered in the neighborhood, the value represents the expected number of salient points to be included in the neighborhood, and A represents the total area of ​​the image, calculated from the image's height and width. This represents the set of all clusters. The DBSCAN algorithm divides the input point set into multiple clusters according to density. This represents the coordinates of the anchor point of the current cluster, i.e., the geometric center of all points in the cluster. x and y are the x and y coordinates of the current anchor point, and n represents the number of significant points in the current cluster.

[0069] By using the Frobenius norm to calculate the pixel gradient changes between frames, we can accurately screen preliminary candidates for significant regions and optimize computational efficiency. By providing a global-scale quantitative basis through the changes in pixel values ​​between frames, this method avoids the computational complexity of the traditional pixel-by-pixel comparison method and can more sensitively capture the changing trend of the entire frame.

[0070] By using Gaussian blur and Sobel edge extraction, irrelevant noise is removed and salient regions are highlighted, improving the purity of edge characteristic data. By calculating the Euclidean distance matrix of salient region points, spatial features are made hierarchically structured and globally consistent. By combining the calculation of local and global neighborhood radii, the spatial distribution scale of the current salient data is dynamically matched. The global neighborhood radius is dynamically defined based on the overall image area and salient point density information, ensuring that the clustering algorithm is adaptable to images of different scales. Whether it is a sparse target distribution scene in a large space or a dense target distribution scene in a small space, this bidirectional calculation of the neighborhood radius ensures efficient extraction and analysis.

[0071] The density-based clustering algorithm DBSCAN is used to extract salient points in a concentrated manner, ensuring the logical uniqueness and local integrity of the extracted results. This solves the problem of traditional k-Means clustering's heavy reliance on initial cluster partitioning in densely varied scenarios, while avoiding errors that may arise from random initialization. Combined with the dynamic calculation of neighborhood radii, the resulting salient clusters not only possess spatial concentration but also eliminate the local boundary discontinuity problem caused by the staircase effect. Anchor point values ​​are generated by calculating the geometric centers of salient region points, providing a precise foundation for efficient localization and behavioral feature extraction in behavior recognition. Subsequent behavioral feature extraction and target region optimization are directly based on the anchor point positions, achieving effective constraints on salient target regions.

[0072] S2 defines a local segmentation window for each anchor point coordinate, forming a local window set. A shallow convolutional network is used to generate feature maps for local feature extraction. Figure 2 The binary mask is calculated by quantization, the vertex coordinates of the rectangular bounding box are defined, the rectangular bounding box is determined, and the rotation IoU loss function is used as the evaluation index of the bounding box fitting quality to dynamically optimize the rectangular bounding box.

[0073] Preferably, the vertex coordinates of the rectangular bounding box are defined, the rectangular bounding box is determined, and the rotation IoU loss function is used as an evaluation index for the quality of the bounding box fitting. The rectangular bounding box is then dynamically optimized, including...

[0074] Based on the set of anchor points, a local segmentation window is defined for the coordinates of each anchor point, resulting in local image patches centered on the anchor points, which are then combined into a set of local windows, represented as follows:

[0075]

[0076]

[0077]

[0078] in, and These represent the horizontal and vertical half-widths of the local window, respectively, and m represents the number of anchor points. and These represent the specific coordinates of the anchor points. and Let represent the mean coordinates of the anchor point set, Indicates anchor point ( , A local image patch centered on )

[0079] For each local image patch, a shallow convolutional network is used to generate a feature map, which is then normalized and represented as follows:

[0080]

[0081] in, This indicates a shallow convolution operation. Represents a local image patch Shallow features;

[0082] Based on the local image patches corresponding to the normalized shallow features, perform local feature extraction. Figure 2 The binary mask is calculated using a value-based method. The maximum and minimum pixel coordinates are determined, and the vertex coordinates of the rectangular bounding box are defined. The rectangular bounding box is then represented as follows:

[0083]

[0084] in, This represents a binary mask representing the salient region of the target at the i-th anchor point. The significance threshold is represented by... The average value (mean of all non-zero elements) is calculated and determined;

[0085] The rotation IoU loss function is used as an evaluation metric for the quality of bounding box fitting. The rectangular bounding box is dynamically optimized, and the key parameters (center point, width, height, and rotation angle) of the bounding box are gradually adjusted using gradient descent. This is represented as:

[0086]

[0087]

[0088] in, Indicates rotational IoU loss. This represents the optimized bounding box. This indicates the current preliminary bounding box. This represents the area of ​​bounding box B. and These represent the parameter vectors in the bounding boxes at updates t+1 and t, respectively. Indicates the learning rate. Indicates the rotational IoU loss with respect to parameters The gradient.

[0089] By defining a local segmentation window using a set of anchor points, the local segmentation window centered on the anchor points directly performs spatial localization on the salient region, which avoids global redundancy in feature extraction of the target region, while preserving the correlation characteristics information within the region. When defining the window size, the mean coordinates of the anchor point set are combined with the statistical distribution to automatically match the saliency distribution scale of the region, thereby improving the robustness and adaptability of the window range definition.

[0090] By using shallow convolutional networks to extract local feature maps, the model effectively captures target texture and edge information, improves regional contrast, extracts edge characteristics and local texture information of target regions, and enables the model to characterize the geometric and appearance characteristics of salient targets. By generating preliminary rectangular bounding boxes based on binary masks, the calculation of salient region boundaries is efficient and convenient, and the image feature map is quickly divided into target region and background region, which greatly simplifies the complexity of subsequent bounding box detection.

[0091] By calculating a dynamic segmentation threshold based on the saliency mean and non-zero elements, the segmentation adaptability and region integrity are enhanced. The global mean of non-zero elements is used as the basis to dynamically adjust the saliency segmentation threshold, so that the threshold definition no longer depends on fixed empirical values ​​and can adapt to the range of changes in the feature matrix in different scenarios.

[0092] By optimizing the bounding box through the rotation IoU loss function, the fitting accuracy of significant region boundaries is improved. This not only considers the shape and position of the target region, but also provides higher accuracy for fitting tilted target regions and non-standard regions by rotating the angle.

[0093] S3, calculate the attention weight map of the salient region, perform enhancement calculation on the channel weights, generate a set of channel-level enhanced feature maps, predict the classification label corresponding to the highest probability through the pre-trained feature classification model, perform gradient calculation on the feature map according to the loss function, reallocate the channel-level attention weights, and update the channel weights.

[0094] Preferably, the attention weight map of the salient region is calculated, and the channel weights are enhanced to generate a set of channel-level enhanced feature maps, including:

[0095] For dynamically optimized rectangular bounding boxes, region images are cut from the original video frames, attention weight maps of salient regions are calculated, and target features of salient regions are optimized through an attention weight allocation mechanism to generate pixel-level salient attention feature maps, represented as follows:

[0096]

[0097]

[0098]

[0099] in, Indicates The i-th cut-out region at the center is designated as the salient region. and This represents the center coordinates of the optimized bounding box. and This represents the width and height of the optimized bounding box. Indicates the location of the i-th salient region Attention weights This represents the normalized shallow features within the i-th salient region. This indicates that within the i-th salient region, for Note that the pixel values ​​at the coordinates should be summed, where , The range of the vertical axis is arrive , This indicates that within the i-th cut region block, for Note that the pixel values ​​at the coordinates should be summed, where, , The range of the horizontal axis is arrive , The coordinate variable representing the pixel position being summed. Represent the target features of the i-th cut region block and form a pixel-level salient attention feature map;

[0100] For each salient attention feature map, global average pooling is performed on each channel to generate statistical features of the channel positions. The sigmoid activation function is then used to generate attention weights for each channel. Enhancement calculations are performed on these channel weights to generate a set of channel-level enhanced feature maps, represented as follows:

[0101]

[0102]

[0103]

[0104] in, H represents the channel-level global information value of the i-th salient region, and H and W represent the width and height of each frame of image data. Indicates the channel index. This represents the channel attention weight for the i-th salient region. This represents the Sigmoid function. and The linear model weights and biases, representing the channel attention weights, are determined through the training process. The channel-level enhancement features of the i-th salient region are represented and formed into a set of enhancement feature maps.

[0105] By dynamically optimizing the rectangular bounding box to cut the image region, the feature separation of salient target regions is improved, ensuring more accurate spatial coverage of salient regions and avoiding interference from background information in feature extraction of target regions. At the same time, key pixel data within the region is preserved. An attention weighting mechanism optimizes the target features of salient regions, strengthening the importance of local target regions and compressing the weight of secondary background information, calculating the saliency contribution of target pixels within a specific region. During attention weighting of each salient region, pixels with high specificity are dynamically boosted while the influence of noise or irrelevant pixels is weakened. This avoids feature redundancy caused by uniform distribution of salient features, making the target features of the region more focused, while reducing the blurring interference of background noise on the overall model.

[0106] By using global average pooling by channel, the channel-level statistical representation capability of features is significantly improved. By generating channel attention weights through Sigmoid activation, the feature participation weights of different channels are adaptively adjusted, improving the channel feature focus and expression sparsity. By enhancing the calculation of channel attention weights, the separation between target regions and distinguishing features is significantly improved, while enhancing the feature expression capability in multi-class and multi-target scenarios. By generating a set of enhanced feature maps at the channel level, multi-scale global and local fusion of salient region features is achieved, improving the robustness and adaptability of content features.

[0107] Furthermore, a set of channel-level enhanced feature maps is generated. The classification label corresponding to the highest probability is predicted using a pre-trained feature classification model. Gradient calculations are performed on the feature maps based on the loss function, channel-level attention weights are reallocated, and the channel weights are updated, including...

[0108] By inputting the channel-level enhanced feature map into the classifier ANN artificial neural network through a pre-trained feature classification model, the classification label corresponding to the highest predicted probability is obtained. The supervision signal loss is calculated based on the true label and the predicted probability, with the loss function being the cross-entropy loss. The gradient of the feature map is calculated according to the loss function, as follows:

[0109]

[0110]

[0111]

[0112] in, Represents the i-th salient region. Indicates the calculation of loss. Indicates the number of significant regions. Indicates the total number of categories. Indicates category index, This represents the one-hot encoding of the actual label, where This represents the actual behavior label for region i, if ,but Otherwise, it is 0. This represents the predicted probability of the ka-th class in the class probability distribution. This represents the update gradient of the feature map. The learning rate for gradient calculation is determined through sample size optimization and is used to control the update magnitude.

[0113] Based on the gradient update results, channel-level attention weights are reallocated and the channel weights are updated. Using the updated channel weights, the enhanced feature maps are recalculated to generate the final optimized feature set, represented as:

[0114]

[0115]

[0116]

[0117] in, This represents the change in attention weight. This represents the gradient of the feature map update for the i-th salient region. This indicates an update to the channel weights. This indicates that the enhanced features are updated and reassembled into an enhanced feature map.

[0118] By using a pre-trained feature classification model to classify channel-level enhanced feature maps for behavior classification, the specificity of feature expression is improved and the behavior category distinction of significant targets is achieved. Compared with directly classifying the original data or preliminary feature maps, using the enhanced feature maps provides more accurate data input and significantly improves the accuracy of classification results. The classification results are supervised and optimized by calculating cross-entropy loss based on real labels, which effectively converges the feature weight distribution of significant regions.

[0119] By dynamically adjusting the feature maps of salient targets through gradient calculation, the characteristic responses of target regions are precisely reshaped. This avoids the limitations of directly setting weights. Instead, by designing a normalized learning rate based on the sample size, each salient target region can be targeted for feature enhancement or suppression based on the actual loss magnitude. The channel-level attention weights are redistributed through gradient results, which deepens the channel optimization of the feature maps of salient regions and strengthens the weight distribution of behavior categories. This not only further refines the feature hierarchy structure of each salient region, but also ensures the focus of channel-level features on truly salient targets in complex scenes. At the same time, it weakens irrelevant information or unnecessary responses of background channels, achieving a more efficient weight optimization goal.

[0120] Channel-level enhanced feature maps not only carry detailed characteristics and behavioral category information of salient regions, but also control the weight balance between features through supervised optimization depth. The final feature set forms highly expressive feature data that combines category characteristics with subtle information within salient regions. It can be directly used as the core data support for anomaly detection, multi-target behavior classification or scene analysis models, thereby significantly improving the accuracy and expressiveness of subsequent tasks.

[0121] S4. Extract neighborhood features of the target behavior region, calculate the variance of channel features to generate abnormal feature responses, and analyze abnormal targets.

[0122] Preferably, the neighborhood features of the target behavior region are extracted, and the variance of the channel features is calculated to generate anomaly feature responses, including:

[0123] From the final optimized feature set and the dynamically optimized rectangular bounding box, neighborhood features of the target behavior region are extracted, and the variance of the channel features is calculated to generate anomaly feature responses, represented as follows:

[0124]

[0125]

[0126] in, This represents the feature set of the target region consisting of all c-channel feature values ​​within the bounding box. This represents the i-th dynamically optimized rectangular bounding box. The channel average value represents the characteristics within the region. This represents the abnormal characteristic response on channel c within the region boundary.

[0127] By extracting neighborhood features of the target behavior region from the final optimized feature set and dynamically optimized rectangular bounding boxes, the spatial and behavioral correlation accuracy of feature extraction is improved. By calculating the channel variance of neighborhood features, the local fluctuations of small anomalies in the feature space are captured, and the sensitivity to target anomalies is improved. By using the dynamically optimized rectangular bounding boxes for variance calculation, the attention to local feature changes is enhanced and the influence of background interference is reduced. By introducing the channel variance within the region, the uniformity of feature distribution is fully quantified, and the sensitivity to the activation characteristics of abnormal targets is improved. By utilizing the interaction between the feature set and the bounding box, the spatial constraints of feature expression and the enhancement of statistical information are combined.

[0128] Further analysis of anomalous targets, including:

[0129] The anomaly response threshold is determined by the sum of the mean and twice the standard deviation of historical anomaly response features. If the anomaly response feature is greater than or equal to the anomaly response threshold, it is judged as an anomaly target.

[0130] This embodiment also provides an analysis system based on AI video recognition behavior monitoring, including,

[0131] The video frame preprocessing module acquires video data, calculates the overall pixel gradient change between adjacent frames, and marks preliminary candidate frames with significant changes.

[0132] The salient region extraction module calculates the edge saliency intensity of each frame, analyzes the salient region point set, and extracts the geometric center of the salient point cluster through clustering to generate an anchor point set;

[0133] The target region boundary optimization module defines a local segmentation window based on anchor points and generates a rectangular bounding box. It then dynamically optimizes the spatial fit of the bounding box using a rotation IoU loss function.

[0134] The salient feature optimization module calculates pixel attention maps of salient regions and enhances channel weights to generate channel-level salient enhanced feature maps.

[0135] The supervised optimization module predicts classification labels using a pre-trained classification model, calculates the gradient of the supervision signal, and reallocates channel weights to optimize the feature map.

[0136] The anomaly detection module extracts neighborhood features of the target behavior region, calculates the variance of the channel features to generate anomaly response, and completes the analysis of the anomaly target.

[0137] This embodiment also provides a computer device applicable to the analysis method based on AI video recognition behavior monitoring, 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 analysis method based on AI video recognition behavior monitoring as proposed in the above embodiment.

[0138] 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.

[0139] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the analysis method for AI video recognition behavior monitoring as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0140] In summary, this invention achieves centralized extraction of salient points through the density-based clustering algorithm DBSCAN, ensuring the logical uniqueness and local integrity of the extracted results. By using a shallow convolutional network to extract local feature maps, it effectively captures target texture and edge information. The bounding box is optimized through a rotation IoU loss function, improving the fitting accuracy of salient region boundaries. The generation of a channel-level enhanced feature map set achieves multi-scale global and local fusion of salient region features, improving the robustness and adaptability of content features. Through gradient calculation, the salient target feature map is dynamically adjusted to accurately reshape the characteristic response of the target region, avoiding the limitations of directly setting weights. Finally, the feature set forms highly expressive feature data that combines category characteristics with subtle information within salient regions.

[0141] 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. An analysis method based on AI video recognition and behavior monitoring, characterized in that, include: Collect video data to calculate the overall pixel gradient change of adjacent frames, mark preliminary candidate frames to calculate the edge saliency intensity of each frame, analyze the set of salient region points, perform clustering to calculate the geometric center point of each cluster as the anchor point value, and form an anchor point set through the center point of each cluster. For each anchor point coordinate, a local segmentation window is defined to form a local window set. A shallow convolutional network is used to generate a feature map. The local feature map is binarized to calculate a binary mask. The vertex coordinates of the rectangular bounding box are defined to determine the rectangular bounding box. The rotation IoU loss function is used as an evaluation index for the quality of the bounding box fitting to dynamically optimize the rectangular bounding box. Calculate the attention weight map of the salient region, perform enhancement calculation on the channel weights, generate a set of channel-level enhanced feature maps, predict the classification label corresponding to the highest probability through a pre-trained feature classification model, perform gradient calculation on the feature map according to the loss function, reallocate the channel-level attention weights, and update the channel weights. Extract neighborhood features of the target behavior region, calculate the variance of channel features to generate abnormal feature responses, and analyze abnormal targets; The geometric center point of each cluster is used as an anchor point value in the clustering calculation. An anchor point set is formed using the center points of each cluster, including... Video data is acquired, and consecutive frames of video image data are combined into an image sequence. The overall pixel gradient change between adjacent frames is calculated using the Frobenius norm. Based on historical experience, a preliminary screening threshold is determined, and frames with a global pixel change rate greater than or equal to the preliminary screening threshold are marked as preliminary candidate frames. For the initial candidate frame set, Gaussian blur is applied for denoising, and the Sobel operator is used to calculate the edge saliency intensity and edge saliency matrix of each frame; Based on the sum of the mean and twice the standard deviation of the edge saliency intensity in historical experience as the region threshold, pixels with edge saliency intensity greater than or equal to the region threshold are regarded as the set of salient region points. The density-based spatial clustering algorithm DBSCAN clusters salient region points, calculates the Euclidean distance between salient region points, determines the average neighborhood distance as the neighborhood radius for each salient region point, calculates the minimum density number of points based on the total number of salient region points, and then calculates the geometric center point of each cluster as the anchor point value. An anchor point set is formed by the center points of each cluster.

2. The analysis method based on AI video recognition behavior monitoring as described in claim 1, characterized in that: The vertex coordinates of the defined rectangular bounding box are used to determine the bounding box. The rotation IoU loss function is then used as an evaluation metric for the bounding box fitting quality to dynamically optimize the bounding box. include, Based on the set of anchor points, a local segmentation window is defined for the coordinates of each anchor point to obtain local image patches centered on the anchor points, which are then combined into a set of local windows. For each local image patch, a shallow convolutional network is used to generate a feature map, which is then normalized. Based on the local image patches corresponding to the normalized shallow features, perform local feature map binarization to calculate the binary mask, and define the vertex coordinates of the rectangular bounding box by determining the maximum and minimum pixel coordinate positions; The rotation IoU loss function is used as an evaluation index for the quality of bounding box fitting. The rectangular bounding box is dynamically optimized, and the key parameters of the bounding box are gradually adjusted using the gradient descent method.

3. The analysis method based on AI video recognition behavior monitoring as described in claim 2, characterized in that: The attention weight map of the salient region is calculated, and the channel weights are enhanced to generate a set of channel-level enhanced feature maps, including... For dynamically optimized rectangular bounding boxes, regional images are cut from the original video frames, attention weight maps of salient regions are calculated, and target features of salient regions are optimized through attention weight allocation mechanisms to generate pixel-level salient attention feature maps. For each salient attention feature map, global average pooling is performed on each channel to generate statistical features of the channel position. The sigmoid activation function is used to generate attention weights for each channel. Enhancement calculations are performed on the channel weights to generate a set of channel-level enhanced feature maps.

4. The analysis method based on AI video recognition behavior monitoring as described in claim 3, characterized in that: The generated set of channel-level enhanced feature maps is used to predict the classification label corresponding to the highest probability through a pre-trained feature classification model. Gradient calculations are performed on the feature maps based on the loss function, channel-level attention weights are reallocated, and the channel weights are updated. include, By using a pre-trained feature classification model, channel-level enhanced feature maps are input into the classifier ANN artificial neural network to obtain the classification label corresponding to the highest predicted probability. The supervision signal loss is calculated based on the true label and the predicted probability, and the loss function is the cross-entropy loss. The gradient of the feature map is calculated according to the loss function. Based on the gradient update results, channel-level attention weights are reallocated and channel weights are updated. Based on the updated channel weight information, the enhanced feature map is recalculated to generate the final optimized feature set.

5. The analysis method based on AI video recognition behavior monitoring as described in claim 4, characterized in that: The process involves extracting neighborhood features of the target behavior region and calculating the variance of the channel features to generate anomaly feature responses. include, From the final optimized feature set and the dynamically optimized rectangular bounding box, the neighborhood features of the target behavior region are extracted, and the variance of the channel features is calculated to generate the abnormal feature response.

6. The analysis method based on AI video recognition behavior monitoring as described in claim 5, characterized in that: The analyzed abnormal targets include, The anomaly response threshold is determined by the sum of the mean and twice the standard deviation of historical anomaly response features. If the anomaly response feature is greater than or equal to the anomaly response threshold, it is judged as an anomaly target.

7. An analysis system based on AI video recognition behavior monitoring, based on the analysis method based on AI video recognition behavior monitoring according to any one of claims 1 to 6, characterized in that: include, The video frame preprocessing module acquires video data, calculates the overall pixel gradient change between adjacent frames, and marks preliminary candidate frames with significant changes. The salient region extraction module calculates the edge saliency intensity of each frame, analyzes the salient region point set, and extracts the geometric center of the salient point cluster through clustering to generate an anchor point set; The target region boundary optimization module defines a local segmentation window based on anchor points and generates a rectangular bounding box. It then dynamically optimizes the spatial fit of the bounding box using a rotation IoU loss function. The salient feature optimization module calculates pixel attention maps of salient regions and enhances channel weights to generate channel-level salient enhanced feature maps. The supervised optimization module predicts classification labels using a pre-trained classification model, calculates the gradient of the supervision signal, and reallocates channel weights to optimize the feature map. The anomaly detection module extracts neighborhood features of the target behavior region, calculates the variance of the channel features to generate anomaly response, and completes the analysis of the anomaly target.

8. 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 analysis method based on AI video recognition behavior monitoring as described in any one of claims 1 to 6.

9. 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 analysis method based on AI video recognition behavior monitoring as described in any one of claims 1 to 6.

Images