Method for grouping videos of video management platform according to video monitoring objects

Abstract

The invention relates to a method for grouping videos according to a video management platform of a video monitoring object, in particular to the field of video monitoring, which comprises the following steps of: acquiring and zooming video monitoring image data under different scenes, constructing a convolutional neural network to output a target feature map, marking a corresponding category label on the category of a target in the video monitoring image data, and marking the category of the target in the video monitoring image data; the method comprises the following steps: segmenting a target area and a scene area of video monitoring image data, constructing a decision tree model, inputting segmented image data of the scene area to classify each scene area, calculating position similarity between monitoring cameras by using an Euclidean distance, and calculating the position similarity between the monitoring cameras by using the Euclidean distance; video monitoring image data with the same monitoring scene and similar positions are divided into the same video stream group through a K-means clustering algorithm, entity nodes and relation edges are combined, and a complete scene knowledge graph is constructed to verify target classification and scene video stream group classification of the video monitoring image data; and the accuracy of video classification and grouping is further verified.

Description

Technical Field

[0001] The present invention relates to the field of video monitoring, and more specifically, to a method for grouping videos on a video management platform according to video monitoring objects. Background Art

[0002] With the continuous development of information technology in the transportation industry, the improvement of infrastructure, the wider coverage of surveillance cameras, and the exponential growth of video data, the huge scale of video data is very challenging for manual processing and analysis. In practical applications, especially in the monitoring of infrastructure such as roads and bridges, the demand for automatic classification and grouping of videos related to a certain infrastructure is increasing. Therefore, it is very necessary to further optimize and improve the method of automatic classification and grouping of camera videos according to video surveillance objects.

[0003] In view of the above situation, there is an urgent need for a method for grouping videos on a video management platform according to video surveillance objects. In the case of traditional manual video classification and grouping, it can further assist managers to automatically group camera video resources according to video surveillance objects, so as to facilitate the rapid formation of video groups during real-time monitoring of an object such as roads and bridges, effectively reduce the time of video search and switching, and provide convenience for managers. Summary of the invention

[0004] The present invention aims at the technical problems existing in the prior art and provides a method for grouping videos on a video management platform according to video surveillance objects to solve the problems raised in the above-mentioned background technology.

[0005] The technical solution of the present invention to solve the above technical problem is as follows: A method for grouping videos on a video management platform according to video surveillance objects comprises the following steps:

[0006] S101. Collect and scale video surveillance image data in different scenarios, construct a convolutional neural network to output a target feature map of the video surveillance image data, input a region proposal network to use a sliding window to generate multiple anchor boxes of different sizes and proportions as candidate target bounding boxes, output the category of the target in the video surveillance image data and label it with the corresponding category label;

[0007] S102. Segment the target area and the scene area of ​​the video surveillance image data and distinguish them, construct a decision tree model and input the segmented image data of the scene area to classify each scene area;

[0008] S103. Obtain the location information data of the surveillance camera and convert it into a feature vector representation, use the Euclidean distance to calculate the location similarity between the surveillance cameras, and use the K-means clustering algorithm to divide the video surveillance image data of the same surveillance scene and similar location into the same video stream group;

[0009] S104. Construct the segmented image data of the scene area into entity nodes and relationship edges of the knowledge graph, combine the entity nodes and relationship edges and construct a complete scene knowledge graph to verify the target classification of the video surveillance image data and the scene video stream group classification;

[0010] In a preferred embodiment, in S101, a surveillance camera is used to collect video surveillance image data in different scenes, the scaling target size of the video surveillance image data is fixed, the position of each pixel in the video surveillance image data is traversed, the nearest pixel position corresponding to each pixel in the video surveillance image data is found and assigned and filled, and the traversal, search and assignment and filling steps are repeated until the complete video surveillance image data is traversed.

[0011] Furthermore, a convolutional neural network is constructed and the video surveillance image data is received through the input layer, a convolutional layer is added and a batch activation function layer is activated, ReLU is used as the activation function layer to introduce nonlinearity and a convolution kernel is used to perform a convolution operation on the target in the video surveillance image data. The specific formula of the convolution operation is:

[0012]

[0013] Wherein S(i,j) represents an element of the target feature in the video surveillance image data, I(m,n) represents the pixel value of the target feature in the video surveillance image data, K(im,jn) represents the weight value of the target feature in the video surveillance image data, m and n represent the row and column indexes inside the convolution kernel respectively, i and j represent the row and column indexes of the target feature in the video surveillance image data respectively, and the target features in the video surveillance image data output by the convolution layer are downsampled through the pooling layer, and the convolution and pooling operations are repeated. The target features in the video surveillance image data are flattened into a one-dimensional vector through the fully connected layer and input into the fully connected layer. The fully connected layer is used to map the target features in the video surveillance image data through the weight matrix and output the target feature map of the video surveillance image data.

[0014] Furthermore, the target feature map of the video surveillance image data is used as the input of the region proposal network. A sliding window is used to start from the upper left corner of the feature map and move to the lower right with the step size. Multiple Anchor Boxes of different sizes and proportions are generated at each sliding window position as candidate target bounding boxes. A convolution operation is performed on the target feature map at each sliding window position to extract the target area features. Each Anchor Box is classified through a small convolutional neural network to determine whether the target area features contain the target, and the position, size and proportion of the Anchor Box are adjusted. The generated candidate target bounding boxes are screened and deduplicated through non-maximum suppression, and the target candidate boxes with high confidence are retained. The category of the target in the video surveillance image data is output and marked with the corresponding category label.

[0015] In a preferred embodiment, in S102, the target area and the scene area of ​​the video surveillance image data are segmented and distinguished, a pixel point located at the center point inside the target area is selected as a seed point, a growth criterion is defined based on the similarity between the pixels, the grayscale value between the pixels is used as the growth criterion, and when the difference between the grayscale value of the pixel point and the grayscale value of the seed point in the target area is less than a set threshold, the pixel point is added to the target area, the color difference between the pixels is used as the growth criterion, and when the difference between the color value of the pixel point and the color value of the seed point in the target area is less than a set difference threshold, the pixel point is added to the target area, the seed point is used as the starting point of the target area, and the neighborhood pixels of the seed point are traversed. When the pixel point meets the growth criterion and has not yet been added to the target area, it is added to the target area and continues to grow as a new seed point, and the growth is continuously iterated until there are no pixels that meet the growth criterion, and finally the segmented image data representing the target area and the scene area is obtained.

[0016] Furthermore, a decision tree model is constructed and the segmented image data of the scene area is input, the color distribution features of the segmented image data of the scene area are extracted using a color histogram, corner point and line segment features are extracted from the segmented image data of the scene area through edge detection, the color distribution features and corner point and line segment features of the segmented image data of the scene area are selected as the division features of the current node and the division points are determined, the color distribution features and corner point and line segment features of the segmented image data of the scene area are divided into multiple monitoring scene feature subsets according to the division features and the division points, wherein each monitoring scene feature subset corresponds to the segmented image data of a different scene area, the division steps are recursively repeated to establish a subtree until the maximum depth of the decision tree is reached, and the current node is marked as a leaf node and the classification value of the leaf node is obtained.

[0017] In a preferred embodiment, in S103, the surveillance camera API is connected to obtain the location information data of the surveillance camera and convert it into a feature vector representation, and the location similarity between the surveillance cameras is calculated using the Euclidean distance combined with the feature vector representation of the location information data. The specific formula is:

[0018]

[0019] Among them, d(p,q) represents the Euclidean distance between two surveillance cameras, p i represents the feature vector representation of the first surveillance camera p in the i-th dimension, q i represents the feature vector representation of the second surveillance camera q in the i-th dimension, and n represents the total number of dimensions for comparing the feature vectors of the location information data of the two surveillance cameras.

[0020] Furthermore, K initial cluster centers are randomly selected through the K-means clustering algorithm, and the Euclidean distance and the classification value of the segmented image data of the scene area are assigned to the cluster center closest to it. The cluster center is updated and the average value of each cluster is used as the new cluster center. The iterative assignment and update operations are repeated until the cluster center no longer changes or the maximum number of iterations is reached. The K-means clustering algorithm is used to extract and group the video surveillance image data of the same monitoring scene and similar location into the same video stream group according to the similarity.

[0021] In a preferred embodiment, in S104, the segmented image data of the scene area is constructed into entity nodes of the knowledge graph, and the relationship edges of the knowledge graph are constructed, the entity nodes and the relationship edges are combined and a complete scene knowledge graph is constructed to verify the target classification of the video surveillance image data and the scene video stream group classification, the actual nodes of the scene area of ​​the video surveillance image data are obtained by using the knowledge graph query language, and the actual nodes are matched and compared with the entity nodes of the knowledge graph, and the association between the nodes is understood by comparing the attribute information of the actual nodes with the attribute information of the entity nodes of the knowledge graph.

[0022] The beneficial effects of the present invention are: with the help of region proposal network and sliding window technology, multiple candidate target bounding boxes are generated, and the target area features are extracted through convolution operation. Using a small CNN to classify each candidate target bounding box, determine whether it contains the target, and adjust the position and size of the bounding box, it is helpful to accurately detect the target in the monitoring image data and output the corresponding category label, reducing the possibility of manual participation and human error, and improving the accuracy and robustness of target detection and reducing false detection results by screening and deduplicating candidate target boxes through non-maximum suppression. The target and background in the monitoring image are effectively distinguished through target area and scene area segmentation, improving the accuracy and reliability of the segmentation results, and using color histogram and edge detection to extract the characteristics of the scene area and build a decision tree model to realize automatic recognition and classification of different scenes. The Euclidean distance is used to calculate the position similarity between surveillance cameras, and the degree of position difference between surveillance cameras is objectively evaluated. The K-means clustering algorithm is used to effectively organize and manage data in the same monitoring scenario. Setting a similarity warning threshold helps to improve the sensitivity and timeliness of the monitoring system to abnormal situations, ensure the accuracy and reliability of monitoring data, and introduce knowledge graphs to determine entity similarity, making the grouping process more intelligent. Knowledge graphs can capture the deep connections between entities, thereby more accurately judging the similarity between videos, helping to discover potential errors or inconsistencies in classification results, and improving the accuracy and credibility of monitoring data classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the method of the present invention; DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0025] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, "plurality" means two or more, unless otherwise clearly and specifically defined.

[0026] In the description of the present application, the term "for example" is used to mean "used as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any technician in the field to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes will not be elaborated in detail to avoid unnecessary details to obscure the description of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed in the present application.

[0027] This embodiment provides Figure 1 A method for grouping videos on a video management platform according to video surveillance objects is shown, specifically comprising the following steps:

[0028] S101. Collect and scale video surveillance image data in different scenarios, construct a convolutional neural network to output a target feature map of the video surveillance image data, input a region proposal network to use a sliding window to generate multiple anchor boxes of different sizes and proportions as candidate target bounding boxes, output the category of the target in the video surveillance image data and label it with the corresponding category label;

[0029] Furthermore, surveillance cameras are used to collect video surveillance image data in different scenarios, the scaling target size of the video surveillance image data is fixed, the position of each pixel in the video surveillance image data is traversed, the nearest pixel position corresponding to each pixel in the video surveillance image data is found and assigned and filled, and the traversal, search and assignment and filling steps are repeated until the complete video surveillance image data is traversed. This can realize the scaling operation of the video surveillance image data, change the size of the video surveillance image data, ensure that the scaled video surveillance image data can still better express the information of the original image, and minimize the distortion and blur between pixels.

[0030] Furthermore, a convolutional neural network is constructed and the video surveillance image data is received through the input layer, a convolutional layer is added and a batch activation function layer is activated, ReLU is used as the activation function layer to introduce nonlinearity and a convolution kernel is used to perform a convolution operation on the target in the video surveillance image data. The specific formula of the convolution operation is:

[0031]

[0032] Wherein S(i,j) represents an element of the target feature in the video surveillance image data, I(m,n) represents the pixel value of the target feature in the video surveillance image data, K(im,jn) represents the weight value of the target feature in the video surveillance image data, m and n represent the row and column indexes inside the convolution kernel respectively, i and j represent the row and column indexes of the target feature in the video surveillance image data respectively, and the target features in the video surveillance image data output by the convolution layer are downsampled through the pooling layer, and the convolution and pooling operations are repeated. The target features in the video surveillance image data are flattened into a one-dimensional vector through the fully connected layer and input into the fully connected layer. The fully connected layer is used to map the target features in the video surveillance image data through the weight matrix and output the target feature map of the video surveillance image data, which is used to capture the target feature maps at different scales and angles, reduce the size of the target feature map while retaining important features.

[0033] Furthermore, the target feature map of the video surveillance image data is used as the input of the region proposal network. A sliding window is used to start from the upper left corner of the feature map and move to the lower right with the step size. A plurality of Anchor Boxes of different sizes and proportions are generated at each sliding window position as candidate target bounding boxes, wherein the Anchor Boxes can diversely cover the different shapes and sizes of targets appearing in the video surveillance image data. A convolution operation is performed on the target feature map at each sliding window position to extract the target region features. Each Anchor Box is classified by a small convolutional neural network to determine whether the target region features contain the target, and the position, size and proportion of the Anchor Box are adjusted. The generated candidate target bounding boxes are screened and deduplicated by non-maximum suppression, and the target candidate boxes with high confidence are retained. The categories of the targets in the video surveillance image data are output and marked with corresponding category labels, which are used to cover various shapes and sizes of targets in the video surveillance image data, realize the function of multi-scale target detection, improve the accuracy and robustness of target detection, and reduce the appearance of repeated target boxes.

[0034] S102. Segment the target area and the scene area of ​​the video surveillance image data and distinguish them, construct a decision tree model and input the segmented image data of the scene area to classify each scene area;

[0035] Furthermore, the target area and scene area of ​​the video surveillance image data are segmented and distinguished, and the pixel point located at the center point of the target area is selected as the seed point. The growth criterion is defined according to the similarity between the pixels to determine whether to add the pixel point to the target area. The grayscale value between the pixels is used as the growth criterion. When the difference between the grayscale value of the pixel point and the grayscale value of the seed point in the target area is less than a set threshold, the pixel point is added to the target area, where the threshold is set to 20. The color difference between the pixels is used as the growth criterion. When the difference between the color value of the pixel point and the color value of the seed point in the target area is less than a set difference threshold, the pixel point is added to the target area. The pixel point is added to the target area, where the difference threshold is set to 5, and the seed point is used as the starting point of the target area. The neighborhood pixels of the seed point are traversed. When the pixel point meets the growth criterion and has not yet been added to the target area, it is added to the target area and continues to grow as a new seed point. The growth is continuously iterated until there are no pixels that meet the growth criterion. Finally, segmented image data representing the target area and the scene area is obtained, where the pixel value of the target area is set to a red label, and the pixel value of the scene area is set to a blue label. The target area and the background scene area in the video surveillance image data are effectively segmented and distinguished, thereby improving the accuracy and reliability of the segmentation results.

[0036] Furthermore, a decision tree model is constructed and the segmented image data of the scene area is input, the color distribution features of the segmented image data of the scene area are extracted using a color histogram, the corner point and line segment features of the segmented image data of the scene area are extracted through edge detection, the color distribution features and the corner point and line segment features of the segmented image data of the scene area are selected as the division features of the current node and the division points are determined, the color distribution features and the corner point and line segment features of the segmented image data of the scene area are divided into multiple monitoring scene feature subsets according to the division features and the division points, wherein each monitoring scene feature subset corresponds to the segmented image data of a different scene area, the division steps are recursively repeated to establish a subtree until the maximum depth of the decision tree is reached, and the current node is marked as a leaf node and the classification value of the leaf node is obtained, that is, the classification value of the segmented image data of the current scene area, so as to better understand the characteristics and structure of the monitoring scene and realize automatic recognition and classification of different scenes.

[0037] S103. Obtain the location information data of the surveillance camera and convert it into a feature vector representation, use the Euclidean distance to calculate the location similarity between the surveillance cameras, and use the K-means clustering algorithm to divide the video surveillance image data of the same surveillance scene and similar location into the same video stream group;

[0038] Furthermore, the surveillance camera API is connected to obtain the location information data of the surveillance camera and converted into a feature vector representation. The location similarity between the surveillance cameras is calculated using the Euclidean distance combined with the feature vector representation of the location information data. The specific formula is:

[0039]

[0040] Among them, d(p,q) represents the Euclidean distance between two surveillance cameras, p i represents the feature vector representation of the first surveillance camera p in the i-th dimension, q i Represents the feature vector representation of the second surveillance camera q in the i-th dimension, and n represents the total number of dimensions for comparing the feature vectors of the location information data of the two surveillance cameras. The smaller the Euclidean distance, the greater the location similarity of the two surveillance cameras, that is, the smaller the location difference of the two surveillance cameras. The larger the Euclidean distance, the smaller the location similarity of the two surveillance cameras, that is, the larger the location difference of the two surveillance cameras. Objectively evaluating the location difference between surveillance cameras is helpful to quickly discover surveillance cameras with similar locations.

[0041] Furthermore, K initial cluster centers are randomly selected through the K-means clustering algorithm, and the Euclidean distance and the classification value of the segmented image data of the scene area are assigned to the cluster center closest to it. The cluster center is updated and the average value of each cluster is used as the new cluster center. The iterative assignment and update operations are repeated until the cluster center no longer changes or the maximum number of iterations is reached. The K-means clustering algorithm is used to extract and group the video surveillance image data of the same monitoring scene and similar location into the same video stream group according to the similarity.

[0042] Furthermore, the similarity warning threshold is set to 0.95. When the target classification and the scene video stream group classification exceed the similarity warning threshold, the same category label is marked. When the target classification and the scene video stream group classification do not exceed the similarity warning threshold, an alarm is triggered and a manual intervention signal is prompted for manual inspection. This can promptly detect abnormal monitoring data and ensure the reliability and security of the grouping.

[0043] S104. Construct the segmented image data of the scene area into entity nodes and relationship edges of the knowledge graph, combine the entity nodes and relationship edges and construct a complete scene knowledge graph to verify the target classification of the video surveillance image data and the scene video stream group classification;

[0044] Furthermore, the segmented image data of the scene area are constructed into entity nodes of the knowledge graph, wherein each entity includes scene area attribute information of the video surveillance image data, and relational edges of the knowledge graph are constructed, including the existence of "connection relationships" and "spanning relationships" in the scene area of ​​the video surveillance image data. The entity nodes and relational edges are combined and a complete scene knowledge graph is constructed to verify the target classification and scene video stream group classification of the video surveillance image data. The actual nodes of the scene area of ​​the video surveillance image data are obtained by using the knowledge graph query language to match and compare with the entity nodes of the knowledge graph to find the entity nodes in the knowledge graph corresponding to the scene area of ​​the video surveillance image data. By comparing the attribute information of the actual nodes with the attribute information of the entity nodes of the knowledge graph, the association between the nodes is understood, which helps to analyze the correlation and connection between different areas in the monitoring scene, deeply analyze the association characteristics between different scene areas, help to more accurately understand the background information and characteristics of the video surveillance image data, and improve the accuracy and efficiency of classification.

[0045] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0046] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0047] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0048] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0049] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0050] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0051] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for grouping videos on a video management platform according to video surveillance objects, characterized in that: The specific steps include: S101. Collect and scale video surveillance image data in different scenarios, construct a convolutional neural network to output a target feature map of the video surveillance image data, input a region proposal network to use a sliding window to generate multiple anchor boxes of different sizes and proportions as candidate target bounding boxes, output the category of the target in the video surveillance image data and label it with the corresponding category label; S102. Segment the target area and the scene area of ​​the video surveillance image data and distinguish them, construct a decision tree model and input the segmented image data of the scene area to classify each scene area; S103. Obtain the location information data of the surveillance camera and convert it into a feature vector representation, use the Euclidean distance to calculate the location similarity between the surveillance cameras, and use the K-means clustering algorithm to divide the video surveillance image data of the same surveillance scene and similar location into the same video stream group; S104. Construct the entity nodes and relationship edges of the knowledge graph for the segmented image data of the scene area, combine the entity nodes and relationship edges and construct a complete scene knowledge graph to verify the target classification of the video surveillance image data and the classification of the scene video stream group.

2. The method for grouping videos on a video management platform according to video surveillance objects according to claim 1, characterized in that: In S101, a surveillance camera is used to collect video surveillance image data in different scenes, the scaling target size of the video surveillance image data is fixed, the position of each pixel in the video surveillance image data is traversed, the nearest pixel position corresponding to each pixel in the video surveillance image data is found and assigned and filled.

3. The method for grouping videos on a video management platform according to video surveillance objects according to claim 1, characterized in that: In the S101, the specific steps of constructing a target feature map of the video surveillance image data output by the convolutional neural network are as follows: constructing a convolutional neural network and receiving the video surveillance image data through an input layer, adding a convolutional layer and batch activating the function layer, using ReLU as an activation function layer to introduce nonlinearity and using a convolution kernel to perform a convolution operation on the target in the video surveillance image data, performing a downsampling operation on the target features in the video surveillance image data output by the convolutional layer through a pooling layer, repeating the iterative convolution and pooling operations, flattening the target features in the video surveillance image data into a one-dimensional vector through a fully connected layer, and inputting it into the fully connected layer, using the fully connected layer to map the target features in the video surveillance image data through a weight matrix and outputting a target feature map of the video surveillance image data, using the target feature map of the video surveillance image data as the input of the region proposal network, using a sliding window starting from the upper left corner of the feature map and continuously moving to the lower right with the step size, and generating multiple Anchors of different sizes and proportions at each sliding window position Boxes are used as candidate target bounding boxes, and a convolution operation is performed on the target feature map at each sliding window position to extract the target area features. Each AnchorBoxes is classified through a small convolutional neural network to determine whether the target area features contain the target, and the position, size and scale of the Anchor Box are adjusted. The generated candidate target bounding boxes are screened and deduplicated through non-maximum suppression, and the target candidate boxes with high confidence are retained. The category of the target in the video surveillance image data is output and marked with the corresponding category label.

4. The method for grouping videos on a video management platform according to video surveillance objects according to claim 3, characterized in that: The specific formula of the convolution operation is: Where S(i,j) represents an element of the target feature in the video surveillance image data, I(m,n) represents the pixel value of the target feature in the video surveillance image data, K(im,jn) represents the weight value of the target feature in the video surveillance image data, m and n represent the row and column indexes inside the convolution kernel, respectively, and i and j represent the row and column indexes of the target feature in the video surveillance image data, respectively.

5. The method for grouping videos on a video management platform according to video surveillance objects according to claim 1, characterized in that: In the S102, the target area and the scene area of ​​the video surveillance image data are segmented and distinguished, a pixel point located at the center point of the target area is selected as a seed point, a growth criterion is defined according to the similarity between the pixels, the gray value and the color difference between the pixels are used as the growth criterion, the seed point is used as the starting point of the target area, the neighborhood pixels of the seed point are traversed, and when the pixel point meets the growth criterion and has not yet been added to the target area, it is added to the target area and continues to grow as a new seed point, and the growth is continuously iterated until there is no pixel point that meets the growth criterion, and finally the segmented image data representing the target area and the scene area is obtained, and a decision tree is constructed. A decision tree model is built and the segmented image data of the scene area is input. The color distribution features of the segmented image data of the scene area are extracted using the color histogram. The corner point and line segment features of the segmented image data of the scene area are extracted through edge detection. The color distribution features and the corner point and line segment features of the segmented image data of the scene area are selected as the partition features of the current node and the partition points are determined. The color distribution features and the corner point and line segment features of the segmented image data of the scene area are divided into multiple monitoring scene feature subsets according to the partition features and the partition points. The partitioning steps are recursively repeated to establish a subtree until the maximum depth of the decision tree is reached, and the current node is marked as a leaf node and the classification value of the leaf node is obtained.

6. The method for grouping videos on a video management platform according to video surveillance objects according to claim 1, characterized in that: In S103, the surveillance camera API is connected to obtain the location information data of the surveillance camera and converted into a feature vector representation, the Euclidean distance is combined with the feature vector representation of the location information data to calculate the location similarity between the surveillance cameras, K initial clustering centers are randomly selected through the K-means clustering algorithm, the Euclidean distance and the classification value of the segmented image data of the scene area are assigned to the cluster center closest to it, the cluster center is updated and the average value of each cluster is used as the new cluster center, the iterative assignment and update operations are repeated until the cluster center no longer changes or the maximum number of iterations is reached, and the K-means clustering algorithm is used to extract and group the video surveillance image data of the same monitoring scene and similar location into the same video stream group according to the similarity.

7. The method for grouping videos on a video management platform according to video surveillance objects according to claim 6, characterized in that: The specific formula for calculating the position similarity between surveillance cameras is: Among them, d(p,q) represents the Euclidean distance between two surveillance cameras, p i represents the feature vector representation of the first surveillance camera p in the i-th dimension, q i represents the feature vector representation of the second surveillance camera q in the i-th dimension, and n represents the total number of dimensions for comparing the feature vectors of the location information data of the two surveillance cameras.

8. The method for grouping videos on a video management platform according to video surveillance objects according to claim 1, characterized in that: In S104, the segmented image data of the scene area is constructed into entity nodes of the knowledge graph, and the relationship edges of the knowledge graph are constructed. The entity nodes and the relationship edges are combined to construct a complete scene knowledge graph to verify the target classification of the video surveillance image data and the scene video stream group classification. The actual nodes of the scene area of ​​the video surveillance image data are obtained by using the knowledge graph query language to match and compare with the entity nodes of the knowledge graph. The association between the nodes is understood by comparing the attribute information of the actual nodes with the attribute information of the entity nodes of the knowledge graph.

Images