Security risk behavior detection method, system, equipment and medium

Abstract

A method, system, device and medium for detecting safety risk behaviors, which obtain an image stream containing a target object; perform foreground-background separation on the image stream to obtain a foreground information image stream; perform node identification on the foreground information image stream to obtain human posture node information; and determine the behavior corresponding to the target object based on the human posture node information. The present invention performs foreground-background separation on the monitoring image stream of electric power operation and maintenance personnel to obtain a foreground information image stream, and then uses the separated foreground information to detect safety risk behaviors of electric power operation and maintenance personnel, thereby improving detection accuracy. Based on the foreground information image stream, the human posture nodes of the electric power operation and maintenance personnel are directionally detected to obtain the human posture node information, and based on the human posture node information, the safety risk behaviors of the electric power operation and maintenance personnel are identified, thereby suppressing the problems of false detection and missed detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of security risk behavior detection, and relates to a behavior detection and analysis method for monitoring videos, and specifically to a security risk behavior detection method, system, device and medium. Background Art

[0002] In terms of detecting safety risk behaviors of power operation and maintenance personnel, currently the most widely used are single-frame-based personnel crossing-border detection and single-frame-based human posture detection. Even when using image streams, most applications simply convert the video stream into a single-frame image and then process it using a single-frame image algorithm. Image stream processing technology has the advantages of single-frame image processing and can reuse inter-frame information between multiple frames. At present, there are many cross-application directions related to image streams, so these cross-application issues have also received extensive attention from experts and scholars.

[0003] At present, when power operation and inspection personnel are performing operations, they generally use fixed cameras for monitoring. The existing image stream processing technology has the following problems:

[0004] 1. For fixed cameras, if the power operation and maintenance personnel overlap with the foreground and background image information of the work site, traditional human posture recognition algorithms such as OpenPose are not robust enough for the above situation, which will interfere with the detection of safety risk behaviors of power operation and maintenance personnel, resulting in unsatisfactory accuracy;

[0005] 2. For fixed cameras, if there is an angle deviation between the camera equipment layout and the work site image, traditional human posture recognition algorithms such as OpenPose are not robust enough for the above situation, which will interfere with the detection of safety risk behaviors of power operation and maintenance personnel and easily cause false detection and missed detection problems. Summary of the invention

[0006] In view of the problems in the prior art, the purpose of the present invention is to provide a security risk behavior detection method, system, device and medium, which has high detection accuracy and overcomes the problems of false detection and missed detection in the prior art.

[0007] To achieve the above object, the technical solution adopted by the present invention is:

[0008] A method for detecting security risk behavior comprises the following steps:

[0009] Get the image stream containing the target object;

[0010] Performing foreground and background separation on the image stream to obtain a foreground information image stream;

[0011] Performing node recognition on the foreground information image stream to obtain human body posture node information;

[0012] According to the human body posture node information, a behavior corresponding to the target object is determined.

[0013] Further, obtaining an image stream containing the target object includes the following steps:

[0014] The monitoring video images of the target object in the same power operation and maintenance scenario are collected, and after preprocessing, an image stream containing the target object is obtained.

[0015] Further, performing node recognition on the foreground information image stream to obtain human body posture node information includes the following steps:

[0016] According to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame;

[0017] Based on the free anchor frame, the target object is rotated to obtain an upright human image;

[0018] According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information.

[0019] Further, according to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame, including the following steps:

[0020] According to the detector, the feature map size of the foreground information image stream is reduced, and then feature extraction is performed to obtain a preliminary feature map;

[0021] The preliminary feature map is divided into a key point feature extraction map and a prediction feature extraction map. The key point feature extraction map is processed by a hole convolutional network structure that maintains the size of the feature map, and the prediction feature extraction map is processed by a residual network structure with an encoding and decoding structure. The processing results of the key point feature extraction map and the processing results of the prediction feature extraction map are merged to perform multi-scale prediction to obtain a prediction result; the prediction result is represented by a direction-adaptive rectangular box to obtain a free anchor box.

[0022] Furthermore, the human body posture node information includes: node information, inter-node association relationship information and human body key point information;

[0023] According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information, including the following steps:

[0024] According to the human key node detector for the upright human image, prediction is performed to obtain node information, information on the association relationship between nodes, and human key point information.

[0025] Further, determining the behavior corresponding to the target object according to the human body posture node information includes the following steps:

[0026] The human body posture node information is input into a pre-trained graph neural network model to determine the behavior corresponding to the target object.

[0027] Furthermore, the process of constructing a pre-trained graph neural network model includes the following steps: constructing a graph of the connection relationships between key nodes of the human body, annotating the graph with the target object's safety risk behavior type as a label to form training data, and using the training data to train the graph neural network model to obtain a pre-trained graph neural network model.

[0028] A security risk behavior detection system, comprising:

[0029] An image stream acquisition unit, used for acquiring an image stream containing a target object;

[0030] A foreground-background separation unit, used for performing foreground-background separation on the image stream to obtain a foreground information image stream;

[0031] A node identification unit, used for performing node identification on the foreground information image stream to obtain human body posture node information;

[0032] The target object corresponding behavior determination unit is used to determine the target object corresponding behavior according to the human body posture node information.

[0033] A computer device comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the computer program is executed by the processor, the above-mentioned security risk behavior detection method is implemented.

[0034] A computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the above-mentioned security risk behavior detection method.

[0035] Compared with the prior art, the present invention has the following technical effects:

[0036] In view of the problem that traditional human posture recognition algorithms are sensitive to the overlap of foreground and background image information, the present invention, based on the idea of ​​"fixed background and changing foreground", separates the foreground-background of the monitoring image stream of power operation and inspection personnel to obtain a foreground information image stream, and then uses the separated foreground information to detect the safety risk behavior of power operation and inspection personnel, removes the interference of the background, and thus improves the detection accuracy. In view of the problem that traditional human posture recognition algorithms are sensitive to the deviation between the layout of the camera equipment and the angle of the image at the work site, the human posture nodes of the power operation and inspection personnel are detected in a directed manner based on the foreground information image stream to obtain the human posture node information. Based on the human posture node information, the safety risk behavior of the power operation and inspection personnel is identified, thereby suppressing the problems of false detection and missed detection.

[0037] Furthermore, according to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame; based on the free anchor frame, the human body target is rotated and transformed to obtain an upright human body image; based on the upright human body image, human body posture nodes are detected to obtain human body posture node information, and then identification is performed to improve detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the safety risk behavior detection method for power operation and maintenance personnel.

[0039] Figure 2 This is a schematic diagram of the safety risk behavior detection results of power operation and maintenance personnel.

[0040] Figure 3 This is a framework diagram of the safety risk behavior detection method for power operation and maintenance personnel.

[0041] In the figure, 1 is an image stream acquisition unit, 2 is a foreground-background separation unit, 3 is a node recognition unit, and 4 is a behavior determination unit corresponding to the target object. DETAILED DESCRIPTION

[0042] The present invention is described in detail below with reference to the accompanying drawings.

[0043] The present invention proposes a method for detecting security risk behaviors based on image streams as follows:

[0044] First, in view of the problem that traditional human posture recognition algorithms are sensitive to the overlapping of foreground and background image information, based on the idea of ​​"fixed background and changing foreground", the foreground and background of the monitoring video of power operation and inspection personnel are separated, and then the separated foreground information is used to detect the safety risk behavior of power operation and inspection personnel, thereby improving the detection accuracy. Secondly, in view of the problem that traditional human posture recognition algorithms are sensitive to the deviation of the angle of camera equipment layout and the image of the work site, a human posture recognition strategy for power operation and inspection personnel based on free anchor frames is proposed to suppress false detection and missed detection problems. Finally, the graph neural network is connected in series after the neural network structure of the human posture recognition of power operation and inspection personnel to realize the identification of safety risk behaviors of power operation and inspection personnel.

[0045] See also Figure 1 and Figure 2 The specific steps of the security risk behavior detection method based on image stream proposed by the present invention are as follows:

[0046] S1: Acquire an image stream containing a target object. In the present invention, the target object is an electric power operation and maintenance personnel.

[0047] Specifically, collect monitoring video images of power operation and maintenance personnel in the same power operation and maintenance scenario Divide it into N images, N≤T, and arrange them in chronological order to obtain the monitoring image stream of power operation and inspection personnel. Among them, N is the number of images extracted in time sequence, T is the number of images actually collected, and n 1 is the horizontal resolution, n 2 is the vertical resolution, X i For the i-th frame image, extracting the image for processing is beneficial to reducing the amount of processed data and speeding up the processing speed.

[0048] S2: performing foreground and background separation on the image stream to obtain a foreground information image stream;

[0049] Specifically, in view of the problem that traditional human posture recognition algorithms are sensitive to the overlap of foreground and background image information, based on the idea of ​​"fixed background and changing foreground", the foreground and background of the power operation and maintenance personnel monitoring image stream are separated. The specific process is: the power operation and maintenance personnel monitoring image stream under the same power operation and maintenance scenario is known. Pull it into a column vector and arrange it in chronological order to get the matrix Since the power operation and inspection personnel monitor the image flow The background information in is fixed, so the matrix D can be considered as the superposition of the background information matrix B and the foreground information matrix F. Since the background changes little, the background information matrix B should satisfy the nuclear norm ||B||* as small as possible, and the values ​​of the background part in the foreground information matrix are all 0, so the foreground information matrix F satisfies the 1-norm ||F|| 1 As small as possible.

[0050] Therefore, the problem of separating the foreground information (personnel) and the background information (scene) of the power operation and inspection personnel monitoring image stream can be expressed as the following optimization problem:

[0051] min B,F ||B||*+λ||F|| 1 st D=B+F

[0052] Where λ is the foreground weight, and its value can be adjusted according to the business situation. Solve the background information matrix B and the foreground information matrix F, and restore the column vector in the foreground information matrix F to the foreground information image flow The problem can be solved by using the Exact Augmented Lagrange Multiplier (Exact ALM) algorithm proposed by Lin, Chen, Ma and others.

[0053] S3: Perform node recognition on the foreground information image stream to obtain human body posture node information;

[0054] Specifically, in order to solve the problem that the traditional human posture recognition algorithm is sensitive to the deviation of the camera equipment layout and the angle of the work site image, the human posture recognition of the power operation and inspection personnel is performed based on the free anchor frame. Conduct safety risk behavior detection on power operation and maintenance personnel.

[0055] S3.1: According to the foreground information image stream, perform human detection based on the free anchor frame to obtain the free anchor frame;

[0056] Design a directional free anchor box detector G for foreground images 1 :F i →P i ,in is the directional detection result of the power operation and inspection personnel in the image, and the directional free anchor box detector G 1 Divided into feature extractor G 1,1 With the multi-task detector G 1,2 Two parts.

[0057] Feature Extractor G 1,1 Using a parallel structure, firstly connect F i The feature map size is reduced by 7x7 convolution, and preliminary feature extraction is completed through a set of 3x3 convolutions to obtain a preliminary feature map. Then, the preliminary feature map is divided into two parts, the key point feature extraction map and the prediction feature extraction map, according to the location of the key points. The key point feature extraction map is processed by a hole convolution network structure that maintains the size of the feature map to obtain the feature map of the key point feature extraction map; the prediction feature extraction map is processed by a residual network structure of the codec structure to obtain the feature map of the prediction feature extraction map. The feature map of the key point feature extraction map and the feature map of the prediction feature extraction map are merged, and then multi-scale prediction is performed to obtain the directional detection results of the power operation and inspection personnel in the image.

[0058] Directional detection results of power operation and inspection personnel in the image Adopt direction-adaptive rectangular box representation to obtain free anchor box P ij ;P ij ={x ij ,y ij ,w ij ,h ij ,θ ij}, i is the frame number, j is the number of a power operation and inspection personnel. ij is the horizontal coordinate of the center point of the detection box, y ij is the vertical coordinate of the center point of the detection box, w ij is the detection box width, h ij is the height of the detection box, θ ij The detection frame is rotated clockwise around the center.

[0059] S3.2: Rotate the target object based on the free anchor frame to obtain an upright human image;

[0060] According to the free anchor box P ij The parameter value is used to perform a rotation transformation G on each detected human target. 2 :F i ,P ij →Q ij , get the upright human body image The image rotation transformation method is: in, is the coordinate after rotation, x 1 is the horizontal coordinate after rotation, y 1 is the vertical coordinate after rotation, is the coordinate before rotation, x 2 is the horizontal coordinate before rotation, y 2 is the ordinate before rotation. Since the coordinate before rotation The value of is not an integer, which is not convenient for indexing pixels in the image. You can directly use the coordinates before rotation The value of The value is rounded up or down to get the coordinates before rotation The 4 pixel coordinates adjacent to the point are used as the coordinates before rotation using linear interpolation of the 4 pixels. The pixel value of the point can be interpolated to obtain a higher quality image.

[0061] S3.3: Detect human body posture nodes according to the upright human body image and obtain human body posture node information;

[0062] Designed for upright human image Q ij Human key node detector G 3 :Q ij →S ij , where S ij is the key point information of the human body. 3 Divided into feature extractor G 31 And key point association prediction G 32 Two parts. Feature extractor G 31 It consists of convolutional layers and pooling layers, using C 3×3 +C 3×3 +P 2×2 +C 3×3 +C 3×3 +P 2×2 +C 3×3 +C 3×3 +C 3×3 +C 3×3 +P 2×2 +C3×3 +C 3×3 Network structure, where C 3×3 represents 3x3 convolution, P 2×2 Represents a 2x2 maximum pooling layer. Key point association prediction G 32 It consists of 6 parts connected in series. The input of the first part is the feature extractor G 31 The input of each of the remaining parts is the output of the previous part and the feature extractor G 31 This structure ensures that each branch can effectively utilize the local features of the image. 32 Each part includes two branches with the same structure, one of which is used to predict node information and the other is used to predict the relationship information between nodes. The two branch structures of part 1 are: 3×3 +C 3×3 +C 3×3 +C 1×1 +C 1×1 The branch structures of the remaining parts (part 2 to part 6) are: C 7×7 +C 7×7 +C 7×7 +C 7×7 +C 7×7 +C 1×1 +C 1×1 , and finally output the human key point information S ij .

[0063] Directed Free Anchor Box Detector G with Undetermined Training Parameters 1 , and obtain the human key node detector G with determined parameters 1 ;

[0064] Human key node detector G with undetermined training parameters 3 , and obtain the human key node detector G determined by the training parameters 3 , thus we get G:=G 1 G 2 G 3 :F i →S i , and then the human posture node information of power operation and inspection personnel can be obtained Among them, l i Represents a single frame image The number of power operation and inspection personnel included in , s represents the number of power operation and inspection personnel's human posture nodes, and 3 represents 3 values, namely: the x-axis coordinate, y-axis coordinate and confidence score of the power operation and inspection personnel's human posture nodes.

[0065] The image flows through the directed free anchor box detector G 1 and the rotation transformation G2 After processing, we get a positive human image, which can be processed using the OpenPose model (i.e., the human key node detector G 3 )Recognize the upright human body image and obtain the human body posture node information.

[0066] S4: Determine the behavior corresponding to the target object according to the human body posture node information.

[0067] Specifically, the graph neural network is connected in series with the human posture recognition neural network structure of the power operation and inspection personnel to realize the identification of the safety risk behavior of the power operation and inspection personnel. The specific process is as follows:

[0068] The human body posture node information includes: node information, inter-node association relationship information and human body key point information;

[0069] The connection relationship between key nodes of the human body is used to form a graph, and the safety risk behavior type of power operation and inspection personnel is used as annotated label to annotate the graph to form training data. The graph neural network model H:S→P is trained using the training data, and the trained model is run to obtain the safety risk behavior type of power operation and inspection personnel. During the operation of the graph neural network model, the graph composed of the connection relationships between the key nodes of the human body is the input of the graph neural network. Given the time series of the graph, the personnel behavior is finally classified according to the safety risk behavior type of the power operation and maintenance personnel to realize the identification of safety risk behavior of the operation and maintenance personnel.

[0070] Among them, A is the total number of people identified in the current image stream, n is the person number, and Z 1 Indicates the type of security risk behavior code, R 1 Indicates the confidence level of the recognition behavior.

[0071] The present invention adopts a foreground-background separation strategy, that is, in view of the problem that the traditional human posture recognition algorithm is sensitive to the overlapping of foreground and background image information, based on the idea of ​​"fixed background and changing foreground", the monitoring image stream of power operation and inspection personnel is separated from the foreground-background to obtain a foreground information image stream, and then the separated foreground information is used to detect the safety risk behavior of power operation and inspection personnel, thereby improving the detection accuracy;

[0072] The present invention adopts a directed human posture recognition strategy, that is, to address the problem that traditional human posture recognition algorithms are sensitive to the deviation of camera equipment layout and on-site image angles, a human posture recognition strategy for power operation and maintenance personnel is developed based on a free anchor frame. That is, according to the foreground information image stream, the human posture nodes of the power operation and maintenance personnel are directed detected to obtain the human posture node information, thereby suppressing false detection and missed detection problems.

[0073] See also Figure 3, a security risk behavior detection system, comprising:

[0074] An image stream acquisition unit 1 is used to acquire an image stream containing a target object;

[0075] A foreground-background separation unit 2 is used to separate the foreground and background of the image stream to obtain a foreground information image stream;

[0076] A node identification unit 3 is used to perform node identification on the foreground information image stream to obtain human body posture node information;

[0077] The target object corresponding behavior determination unit 4 is used to determine the target object corresponding behavior according to the human body posture node information.

[0078] In one embodiment, the image stream acquisition unit 1 is used to:

[0079] The monitoring video images of the target object in the same power operation and maintenance scenario are collected, and after preprocessing, an image stream containing the target object is obtained.

[0080] In one embodiment, the node identification unit 3 is used to:

[0081] According to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame;

[0082] Based on the free anchor frame, the target object is rotated to obtain an upright human image;

[0083] According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information.

[0084] In one embodiment, the node identification unit 3 is used to:

[0085] According to the detector, the feature map size of the foreground information image stream is reduced, and then feature extraction is performed to obtain a preliminary feature map;

[0086] The preliminary feature map is divided into a key point feature extraction map and a prediction feature extraction map. The key point feature extraction map is processed by a hole convolutional network structure that maintains the size of the feature map, and the prediction feature extraction map is processed by a residual network structure with an encoding and decoding structure. The processing results of the key point feature extraction map and the processing results of the prediction feature extraction map are merged to perform multi-scale prediction to obtain a prediction result; the prediction result is represented by a direction-adaptive rectangular box to obtain a free anchor box.

[0087] The human body posture node information includes: node information, inter-node association relationship information and human body key point information;

[0088] The node identification unit 3 is used for:

[0089] According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information, including the following steps:

[0090] According to the human key node detector for the upright human image, prediction is performed to obtain node information, information on the association relationship between nodes, and human key point information.

[0091] In one embodiment, the behavior determination unit 4 corresponding to the target object is used to:

[0092] The human body posture node information is input into a pre-trained graph neural network model to determine the behavior corresponding to the target object.

[0093] A computer device comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the computer program is executed by the processor, the above-mentioned security risk behavior detection method is implemented.

[0094] A computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the above-mentioned security risk behavior detection method.

[0095] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0096] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application may be combined with each other.

[0097] The present invention may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0098] In the present invention, "module", "device", "system" and the like refer to related entities applied to computers, such as hardware, a combination of hardware and software, software or software in execution, etc. In detail, for example, an element can be, but is not limited to, a process, a processor, an object, an executable element, an execution thread, a program and / or a computer running on a processor. In addition, an application or script program running on a server, a server can all be an element. One or more elements can be in an execution process and / or thread, and an element can be localized on a computer and / or distributed between two or more computers, and can be operated by various computer-readable media. An element can also communicate through local and / or remote processes according to a signal with one or more data packets, for example, a signal from a data that interacts with another element in a local system, a distributed system, and / or a network on the Internet through a signal to interact with other systems.

[0099] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include" and "comprise" include not only those elements, but also other elements that are not explicitly listed, or also include elements inherent to such processes, methods, articles or equipment. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the existence of other identical elements in the process, method, article or equipment that includes the elements.

[0100] 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.

[0101] 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 processor, or other programmable data processing device to produce 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.

[0102] 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.

[0103] 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.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting security risk behavior, characterized in that: The following steps are involved: Get the image stream containing the target object; Performing foreground and background separation on the image stream to obtain a foreground information image stream; Performing node recognition on the foreground information image stream to obtain human body posture node information; Determining a behavior corresponding to the target object according to the human body posture node information; Among them, the foreground information image stream Determined by the following process: The problem of separating foreground information from background information of an image stream is expressed as the following optimization problem: in, is the foreground weight, solve the background information matrix and prospect information matrix , the foreground information matrix The column vector in is restored to the foreground information image stream , n1 is the horizontal resolution, n2 is the vertical resolution, N is the number of images extracted in time sequence, is the i-th frame image, Background information matrix and prospect information matrix The superposition of is the nuclear norm; Performing node recognition on the foreground information image stream to obtain human body posture node information includes the following steps: According to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame; Based on the free anchor frame, the target object is rotated to obtain an upright human image; According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information.

2. The security risk behavior detection method according to claim 1, characterized in that: Obtaining an image stream containing a target object includes the following steps: The monitoring video images of the target object in the same power operation and maintenance scenario are collected, and after preprocessing, an image stream containing the target object is obtained.

3. The security risk behavior detection method according to claim 1, characterized in that: According to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame, including the following steps: According to the detector, the feature map size of the foreground information image stream is reduced, and then feature extraction is performed to obtain a preliminary feature map; The preliminary feature map is divided into a key point feature extraction map and a prediction feature extraction map. The key point feature extraction map is processed by a hole convolutional network structure that maintains the size of the feature map, and the prediction feature extraction map is processed by a residual network structure with an encoding and decoding structure. The processing results of the key point feature extraction map and the processing results of the prediction feature extraction map are merged to perform multi-scale prediction to obtain a prediction result; the prediction result is represented by a direction-adaptive rectangular box to obtain a free anchor box.

4. The security risk behavior detection method according to claim 1, characterized in that: The human body posture node information includes: node information, inter-node association relationship information and human body key point information; According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information, including the following steps: According to the human key node detector for the upright human image, prediction is performed to obtain node information, information on the association relationship between nodes, and human key point information.

5. The security risk behavior detection method according to claim 1, characterized in that: Determining the behavior corresponding to the target object according to the human body posture node information includes the following steps: The human body posture node information is input into a pre-trained graph neural network model to determine the behavior corresponding to the target object.

6. The security risk behavior detection method according to claim 5, characterized in that: The process of building a pre-trained graph neural network model includes the following steps: constructing a graph of the key nodes of the human body and the connection relationships between the nodes, annotating the graph with the target object's safety risk behavior type as a label to form training data, and using the training data to train the graph neural network model to obtain a pre-trained graph neural network model.

7. A security risk behavior detection system, characterized in that: include: An image stream acquisition unit, used for acquiring an image stream containing a target object; A foreground-background separation unit, used for performing foreground-background separation on the image stream to obtain a foreground information image stream; A node identification unit, used for performing node identification on the foreground information image stream to obtain human body posture node information; A target object corresponding behavior determination unit, used for determining the behavior corresponding to the target object according to the human body posture node information; Among them, the foreground information image stream Determined by the following process: The problem of separating foreground information from background information of an image stream is expressed as the following optimization problem: in, is the foreground weight, solve the background information matrix and prospect information matrix , the foreground information matrix The column vector in is restored to the foreground information image stream , n1 is the horizontal resolution, n2 is the vertical resolution, N is the number of images extracted in time sequence, is the i-th frame image, Background information matrix and prospect information matrix The superposition of is the nuclear norm; Performing node recognition on the foreground information image stream to obtain human body posture node information includes the following steps: According to the foreground information image stream, human body detection based on the free anchor frame is performed to obtain the free anchor frame; Based on the free anchor frame, the target object is rotated to obtain an upright human image; According to the upright human body image, the human body posture nodes are detected to obtain the human body posture node information.

8. The security risk behavior detection system according to claim 7, characterized in that: The node identification unit is used to: According to the detector, the feature map size of the foreground information image stream is reduced, and then feature extraction is performed to obtain a preliminary feature map; The preliminary feature map is divided into a key point feature extraction map and a prediction feature extraction map. The key point feature extraction map is processed by a hole convolutional network structure that maintains the size of the feature map, and the prediction feature extraction map is processed by a residual network structure with an encoding and decoding structure. The processing results of the key point feature extraction map and the processing results of the prediction feature extraction map are merged to perform multi-scale prediction to obtain a prediction result; the prediction result is represented by a direction-adaptive rectangular box to obtain a free anchor box.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program that can be run on the processor, and when the computer program is executed by the processor, the security risk behavior detection method described in any one of claims 1-6 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the security risk behavior detection method according to any one of claims 1 to 6.

Images