A power distribution room safety monitoring system and monitoring method based on target detection

Abstract

The present invention discloses a power distribution room safety monitoring system and monitoring method based on target detection, and belongs to the technical field of power distribution room monitoring. The method of the present invention specifically includes the following steps: S1, using high-definition cameras indoors and outdoors of the power distribution room to obtain video streams in real time; S2, training a character data set in a local server, using the YoloV4 target detection algorithm to detect characters, and the detected characters are framed with a detection frame; S3, using the tracking technology based on Kalman filtering and Hungarian algorithm on the local server to track the detected characters; S4, using the reverse perspective transformation on the local server to process the monitored video stream, and converting characters in different directions into characters from a bird's-eye view. The present invention uses the YoloV4 target detection algorithm to detect characters in the video stream, and then tracks the characters through Kalman filtering and the Hungarian algorithm, and then combines the reverse perspective transformation to achieve safety monitoring of the power distribution room.

Description

Technical Field

[0001] The present invention belongs to the technical field of power distribution room monitoring, and in particular relates to a power distribution room safety monitoring system and a monitoring method based on target detection. Background Art

[0002] The power distribution room is an important place for power transfer, protection and control. Traditional power distribution rooms generally have a duty room for security supervision, which consumes a lot of manpower. Most of the current power distribution rooms do not require human supervision. Security conditions are remotely monitored by installing cameras indoors and outdoors, but monitoring personnel are required to watch the monitoring screen in real time. This ignores the situation where the monitoring personnel are not focused or are temporarily absent from their posts, which can easily lead to the theft of equipment in the power distribution room. How to implement round-the-clock supervision of the power distribution room when there is no one guarding the power distribution room and timely and accurately control illegal intrusions has become the focus of research. Summary of the invention

[0003] Purpose of the invention: The present invention proposes a distribution room security monitoring system and monitoring method based on target detection. The main purpose is to provide a security supervision idea to prevent illegal intrusion into the distribution room, to solve the problem that the existing distribution room monitoring personnel cannot view the monitoring screen in time or around the clock. The present invention uses the YoloV4 target detection algorithm to detect people in the video stream, and then tracks the people through Kalman filtering and Hungarian algorithm, and then combines the reverse perspective transformation to achieve security monitoring of the distribution room.

[0004] The above purpose is achieved through the following technical solutions:

[0005] A power distribution room safety monitoring system based on target detection, comprising a server module and a supervision module;

[0006] The monitoring module includes a high-definition camera, a smart door lock, a video intercom and an alarm device; the high-definition camera is installed outdoors at a height of 1 meter around the power distribution room, and is installed in the power distribution room at a height of half a meter around the incoming line cabinet, metering cabinet, and PT cabinet equipment; the smart door lock is installed at the door of the power distribution room, and is used to automatically unlock the power distribution room door through a password, face recognition, and remote control, and automatically lock it in the event of an alarm; the video intercom is used to realize remote communication between the monitoring personnel and the personnel entering the power distribution room, and to understand the dynamics of the power distribution room at the first time; the alarm device is used to remind the monitoring personnel to be vigilant;

[0007] The server module includes a local server installed in a power distribution room and a cloud server for remote communication with monitoring personnel. The local server is connected to a camera and a smart door lock.

[0008] A method for performing power distribution room safety monitoring using the power distribution room safety monitoring system based on target detection comprises the following steps:

[0009] S1. Use high-definition cameras indoors and outdoors in the power distribution room to obtain real-time video streams;

[0010] S2. Train a person dataset in the local server, use the YoloV4 target detection algorithm to detect people, and use a detection box to outline the detected people;

[0011] S3, using the tracking technology based on Kalman filter and Hungarian algorithm on the local server to track the detected person;

[0012] S4. The local server uses reverse perspective transformation to process the monitored video stream, and converts the characters in different directions into characters from a bird's-eye view.

[0013] Furthermore, the use of high-definition cameras indoors and outdoors in the distribution room described in step S1 needs to meet the requirements that the camera installed outdoors can clearly monitor the environment 1 meter away from the distribution room; the camera installed indoors can clearly monitor the environment half a meter away from the incoming cabinet, metering cabinet, and PT cabinet.

[0014] Furthermore, in the process of using the YoloV4 target detection algorithm to detect people in step S2, YoloV4 replaces the backbone feature extraction network of YoloV3 with CSPDarknet53 in structure. After the input image passes through CSPDarknet53, YoloV4 will output three scale feature maps of 13×13, 26×26, and 52×52. The feature maps of different scales contain semantic information of different dimensions; the feature map of 13×13 size will enter the spatial pyramid pooling structure. The spatial pyramid pooling structure will superimpose and convolve the new feature map with the feature map, then enter the network and output it to the feature fusion network; for the 26×26 and 52×52 feature maps, they will first be superimposed and convolved, and then a series of similar downsampling and superimposed convolutions will be performed from bottom to top to fully integrate the features of the three feature maps of different scales, and finally output three YOLO detection heads of 13×13, 26×26, and 52×52 respectively; each detection head in the YoloV4 algorithm contains three sets of adjustment parameters for candidate boxes, and each set of adjustment parameters for candidate boxes contains 1 confidence parameter, 4 parameters for adjusting length, width, and coordinate offset, and 20 category parameters; the YoloV4 algorithm adjusts the coordinates, width, and height of the candidate box through these adjustment parameters, and finally generates a detection box.

[0015] Furthermore, the specific method of step S3 is: the Kalman filter predicts the position of the person at time t+1 based on the current measurement value at time t; the Hungarian algorithm checks whether the person in the current frame is the same as the person detected in the previous frame, and if so, it helps to assign a unique ID number to identify the person.

[0016] Furthermore, the specific method of step S4 is:

[0017] First, the original video needs to be transformed in reverse perspective. The Z in the transformation equation represents the height, and Z=0

[0018] The specific transformation process is as follows:

[0019] Let (u,v) be a two-dimensional coordinate point, (X w ,Y w ,Z w ) is the corresponding three-dimensional coordinate point after the reverse perspective transformation, then,

[0020] [uv 1] T =KRT[X w Y w Z w ] T

[0021] In the above equation, R is the rotation matrix, T is the translation matrix, and

[0022]

[0023]

[0024] h is the camera height, θ is the camera tilt angle, that is, the angle between the camera and the ground,

[0025] K is the following equation:

[0026]

[0027] f is the focal length, k is the focal length u and k v are the calibration parameters in the horizontal and vertical directions respectively. x ,c y ) to correct the optical axis of the image.

[0028] When the above process is completed, the output scene of step S3 is converted from a three-dimensional scene to a two-dimensional scene.

[0029] Beneficial effects:

[0030] The present invention uses the YoloV4 target detection algorithm to detect people in the video stream, then tracks the people through Kalman filtering and the Hungarian algorithm, and combines the reverse perspective transformation to achieve security monitoring of the power distribution room. It can achieve all-weather guarding of the power distribution room when there is no one guarding the power distribution room, and timely and accurately control illegal intrusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a flow chart of the power distribution room safety monitoring method based on target detection proposed by the present invention;

[0032] Figure 2 This is the structure diagram of the YoloV4 algorithm. DETAILED DESCRIPTION

[0033] The power distribution room safety monitoring system based on target detection includes a server module and a supervision module;

[0034] The monitoring module includes a high-definition camera, a smart door lock, a video intercom and an alarm device; the high-definition camera is installed outdoors at a height of 1 meter around the power distribution room, and is installed in the power distribution room at a height of half a meter around the incoming line cabinet, metering cabinet, and PT cabinet equipment; the smart door lock is installed at the door of the power distribution room, and is used to automatically unlock the power distribution room door through a password, face recognition, and remote control, and automatically lock it in the event of an alarm; the video intercom is used to realize remote communication between the monitoring personnel and the personnel entering the power distribution room, and to understand the dynamics of the power distribution room at the first time; the alarm device is used to remind the monitoring personnel to be vigilant;

[0035] The server module includes a local server installed in a power distribution room and a cloud server for remote communication with monitoring personnel. The local server is connected to a camera and a smart door lock.

[0036] like Figure 1 As shown, a method for performing power distribution room safety monitoring using the power distribution room safety monitoring system based on target detection comprises the following steps:

[0037] S1. Use high-definition cameras indoors and outdoors in the power distribution room to obtain real-time video streams.

[0038] Specifically, the high-definition cameras are installed outside and inside the power distribution room. The camera installed outside can clearly monitor the environment 1 meter away from the power distribution room; the camera inside the power distribution room can clearly monitor the environment half a meter away from the incoming line cabinet, metering cabinet, PT cabinet and other equipment.

[0039] S2. Train a person dataset in the local server, use the YoloV4 target detection algorithm to detect people, and use a detection box to outline the detected people.

[0040] Specifically, Yolo is a target detection system based on neural networks. Usually Yolo divides the original image into several grids, which generate feature images of the original image. If Yolo detects the center of an object in the original image through the feature map, the corresponding grid around the center can help quickly locate the object and finally complete the detection.

[0041] Furthermore, YoloV4 replaces the backbone feature extraction network of YoloV3 with CSPDarknet53 in structure. Figure 2 As shown in the figure, after the input image passes through CSPDarknet53, YoloV4 will output three scale feature maps of 13×13, 26×26, and 52×52. Feature maps of different scales contain semantic information of different dimensions. The feature map of 13×13 size will enter the SPP (spatial pyramid pooling) structure. SPP will superimpose and convolve the new feature map obtained with the feature map, then enter the network and output it to the feature fusion network PANet (Path Aggregation Network). For the 26×26 and 52×52 feature maps, they will first be superimposed and convolved, and then a series of similar downsampling and superimposed convolutions will be performed from bottom to top to fully integrate the features of the three feature maps of different scales, and finally output three YOLO detection heads of 13×13, 26×26, and 52×52 respectively.

[0042] Each detection head in the YoloV4 algorithm contains three sets of adjustment parameters for candidate boxes. Each set of adjustment parameters for candidate boxes includes 1 confidence parameter, 4 parameters for adjusting length, width, and coordinate offset, and 20 category parameters. The YoloV4 algorithm uses these adjustment parameters to adjust the coordinates, width, and height of the candidate box and finally generates a detection box.

[0043] In this method, in order to detect each pedestrian in the video stream, four categories of "man", "woman", "boy" and "girl" are selected from the Google Open Images dataset for training. The results can be output as detection boxes.

[0044] S3. Use the tracking technology based on Kalman filtering and Hungarian algorithm on the local server to track the detected person.

[0045] Specifically, the tracker is based on Kalman filtering and the Hungarian algorithm. In terms of multi-target tracking, it can find the optimal matching solution for multiple targets in the upper and lower frames. The Kalman filter predicts the position of the person at time t+1 based on the current measurement value at time t. This is an effective way to track people. The Hungarian algorithm is a combinatorial optimization algorithm. It checks whether the person in the current frame is the same as the person detected in the previous frame, and if so, it helps assign a unique ID number to identify the person. Therefore, once a pedestrian is detected by the YoloV4 detection frame, the tracker will follow the human's route, so the YoloV4 detection frame will move as the pedestrian moves.

[0046] S4. The local server uses reverse perspective transformation to process the monitored video stream, and converts the characters in different directions into characters from a bird's-eye view.

[0047] Furthermore, a warning frame is set up every 1m around the power distribution room from a bird's eye view, and a warning frame is set up every half meter around the incoming line cabinet, metering cabinet, PT cabinet and other equipment in the power distribution room.

[0048] Specifically, the original video needs to be processed by reverse perspective transformation first. Z in the transformation equation represents the height, and Z=0 is set. The specific transformation process is as follows:

[0049] Let (u,v) be a two-dimensional coordinate point, (X w ,Y w ,Z w ) is the corresponding three-dimensional coordinate point after the reverse perspective transformation.

[0050] Then, [uv 1] T =KRT[X w Y w Z w ] T

[0051] In the above equation, R is the rotation matrix, T is the translation matrix,

[0052]

[0053]

[0054] h is the camera height, θ is the camera tilt angle, that is, the angle between the camera and the ground,

[0055] K is the following equation:

[0056] f is the focal length, k is the focal length u and k v are the calibration parameters in the horizontal and vertical directions respectively. x ,c y) to correct the optical axis of the image.

[0057] When the above process is completed, the output scene of S3 is converted from a three-dimensional scene to a two-dimensional scene. In the converted image, it is measured that every 10 pixels is equal to 98 centimeters in reality. Therefore, a warning frame is set at 1m (equivalent to 10 pixels) around the power distribution room from a bird's eye view, and a warning frame is set at half a meter (equivalent to 5 pixels) around the incoming line cabinet, metering cabinet, PT cabinet and other equipment in the power distribution room.

[0058] For people detected by the server outside the power distribution room, if they stay in the 1m warning frame around the power distribution room for more than 5 seconds from a bird's eye view; for people detected by the server inside the power distribution room, if they stay in the half-meter warning frame around the incoming line cabinet, metering cabinet, PT cabinet and other equipment for more than the set time from a bird's eye view, the local server will consider that the person is suspected of illegal intrusion, control the smart door lock to lock, and turn on the alarm device through the cloud server to remind the monitoring personnel to be vigilant. At this time, the monitoring personnel can communicate with the on-site personnel through the video intercom as soon as possible to determine whether the door can be opened again.

Claims

1. A method for safety monitoring of a power distribution room based on target detection, characterized in that: A security monitoring system is included, wherein the system includes a server module and a supervision module; The monitoring module includes a high-definition camera, a smart door lock, a video intercom and an alarm device; the high-definition camera is installed outdoors at a height of 1 meter around the power distribution room, and is installed in the power distribution room at a height of half a meter around the incoming line cabinet, metering cabinet, and PT cabinet equipment; the smart door lock is installed at the door of the power distribution room, and is used to automatically unlock the power distribution room door through a password, face recognition, and remote control, and automatically lock it in the event of an alarm; the video intercom is used to realize remote communication between the monitoring personnel and the personnel entering the power distribution room, and to understand the dynamics of the power distribution room at the first time; the alarm device is used to remind the monitoring personnel to be vigilant; The server module includes a local server installed in the power distribution room and a cloud server for remote communication with monitoring personnel, and the local server is connected to the camera and the smart door lock; The security monitoring method includes the following steps: S1. Use high-definition cameras indoors and outdoors in the power distribution room to obtain real-time video streams; S2. Train a person dataset in the local server, use the YoloV4 target detection algorithm to detect people, and use a detection box to outline the detected people; S3, using the tracking technology based on Kalman filter and Hungarian algorithm on the local server to track the detected person; S4, using reverse perspective transformation to process the monitored video stream on the local server, and converting the characters in different directions into characters from a bird's eye view; The specific method is: First, the original video needs to be processed by reverse perspective transformation; Z in the transformation equation represents the height, and Z=0 The specific transformation process is as follows: Let (u,v) be a two-dimensional coordinate point, (X w ,Y w ,Z w ) is the corresponding three-dimensional coordinate point after the reverse perspective transformation, then, [uv 1] T =KRT[Y w Y w Z w ] T In the above equation, R is the rotation matrix, T is the translation matrix, and h is the camera height, θ is the camera tilt angle, that is, the angle between the camera and the ground, K is the following equation: f is the focal length, k is the focal length u and k v are the calibration parameters in the horizontal and vertical directions respectively. x ,c y ) Correct the optical axis of the image; When the above process is completed, the output scene of step S3 is converted from a three-dimensional scene to a two-dimensional scene.

2. The method for safety monitoring of a power distribution room according to claim 1, characterized in that: In the process of using the YoloV4 target detection algorithm to detect people in step S2, YoloV4 replaces the backbone feature extraction network of YoloV3 with CSPDarknet53 in structure; after the input image passes through CSPDarknet53, YoloV4 will output feature maps of three scales: 13×13, 26×26, and 52×52. Feature maps of different scales contain semantic information of different dimensions; the feature map of 13×13 size will enter the spatial pyramid pooling structure; the spatial pyramid pooling structure will superimpose and convolve the new feature map obtained with the feature map, then enter the network, and output it to the feature fusion network; in 2 For the 6×26 and 52×52 feature maps, they are first stacked and convolved, and then a series of similar downsampling and stacked convolutions are performed from bottom to top to fully integrate the features of the three feature maps of different scales, and finally output three YOLO detection heads of 13×13, 26×26, and 52×52 respectively; each detection head in the YoloV4 algorithm contains three sets of adjustment parameters for candidate boxes, and each set of adjustment parameters for candidate boxes contains 1 confidence parameter, 4 parameters for adjusting length, width, and coordinate offset, and 20 category parameters; the YoloV4 algorithm adjusts the coordinates, width, and height of the candidate box through these adjustment parameters, and finally generates a detection box.

3. The method for safety monitoring of a power distribution room according to claim 1, characterized in that: The specific method of step S3 is: the Kalman filter predicts the position of the person at time t+1 based on the current measurement value at time t; the Hungarian algorithm checks whether the person in the current frame is the same as the person detected in the previous frame, and if so, it helps assign a unique ID number to identify the person.

Images