Perimeter invasion target monitoring system
By designing a perimeter intrusion target monitoring system, using the improved YOLOv8 neural network detection model and SE attention mechanism, high-precision identification and real-time alarm of perimeter intrusion targets are achieved, solving the problem of insufficient protection capabilities of traditional perimeter prevention measures, and improving the response speed and protection capabilities of perimeter security.
Patent Information
- Application Number
- CN202510109437.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-23
- Publication Date
- 2025-06-06
AI Technical Summary
Traditional perimeter prevention measures are difficult to effectively prevent intrusion targets from climbing over the wall illegally, and video surveillance cannot effectively alert and handle them as soon as possible, resulting in insufficient protection capabilities.
A perimeter intrusion target monitoring system is designed, including multiple IP cameras, edge computing smart boxes, acousto-optical alarms and monitoring software platforms. Using the improved YOLOv8 neural network detection model and SE attention mechanism, high-precision identification and real-time alarm of perimeter intrusion targets are achieved.
It improves the response speed and protection capabilities of perimeter security, reduces dependence on cloud servers and network bandwidth through edge computing, and ensures the stability and efficiency of the system, especially in areas with poor network environments.
Smart Images
Figure CN120108103A_ABST
Abstract
Description
Technical Field
[0001] The embodiments of the present application relate to the field of security monitoring technology, and in particular to a perimeter intrusion target monitoring system. Background Art
[0002] The perimeter is the first line of defense of the security protection system and the outermost layer of the security protection in depth system. However, with the continuous changes in the perimeter environment, traditional perimeter prevention measures have been unable to effectively prevent intruders from illegally climbing over the fence and other security risks. Even if video surveillance is deployed on the perimeter, it is still impossible to effectively alert and handle violations as soon as they occur, and the protection capability is still insufficient.
[0003] In recent years, with the rapid development of artificial intelligence technology and edge computing technology, the performance of perimeter intrusion monitoring systems has been significantly improved. These technologies have the advantages of low cost, lightweight and high reliability, and can achieve more flexible deployment. However, most of the current perimeter monitoring systems still rely on the data transmission and processing mode between the front-end sensing device and the cloud server. For the perimeter security of the reservoir area located in the mountainous area, this mode requires a higher network bandwidth and it is difficult to ensure stable data transmission, resulting in poor feedback timeliness of the recognition results, which limits the protection effect of the system.
[0004] Nowadays, smart security has moved from the "digital" era to the "intelligent" era. Intelligent video analysis technology transforms traditional manual monitoring methods into intelligent reasoning analysis, which enables the security system to transform from post-tracing to pre-prevention, effectively improving the perimeter's defense capabilities. However, no research team has yet come up with an effective solution for all-weather, efficient, and real-time monitoring and identification of perimeter intrusion targets. Summary of the invention
[0005] In view of this, an embodiment of the present application proposes a perimeter intrusion target monitoring system, which can achieve high-precision perimeter intrusion target identification and alarm in complex scenarios, effectively improving the perimeter security capability.
[0006] In order to achieve the above-mentioned purpose, an embodiment of the present application proposes a perimeter intrusion target monitoring system, which includes: multiple IP cameras, power supply equipment, switches, edge computing smart boxes, sound and light alarms and monitoring software platforms, wherein multiple IP cameras are installed on fixed poles arranged at preset distances around the perimeter of the target area to achieve full coverage of the viewing angle of the perimeter of the target area; the IP cameras are used to collect video data in real time; the power supply equipment is used to provide power support for all IP cameras; a communication connection is established between the switch and each IP camera, which is used to transmit the video data collected by each IP camera to the edge computing smart box and the monitoring software platform through a self-organizing network; the edge computing smart box is used to detect and analyze multiple video data, and send the detection results and alarm information to the sound and light alarm and the monitoring software platform; the sound and light alarm is used to trigger the sound and light alarm after receiving the detection results and alarm information, and emit strong light and sound waves to drive away the perimeter intrusion targets; the monitoring software platform is used to display the real-time monitoring screen corresponding to each IP camera, and after receiving the detection results and alarm information, it focuses on displaying the alarm content and the real-time monitoring screen corresponding to the IP camera that triggered the alarm.
[0007] Optionally, each IP camera transmits the real-time collected video data in the form of RTSP video stream to the switch in accordance with the real-time streaming protocol, and then transmits it to the edge computing smart box and the monitoring software platform; the edge computing smart box sends the detection results and alarm information in the form of JSON files through HTTP requests to the sound and light alarm and the monitoring software platform.
[0008] Optionally, a pre-trained detection model is deployed in the edge computing smart box, and the edge computing smart box adopts a multi-threaded mode to implement full-process algorithm reasoning for multiple RTSP video streams; assuming that the number of IP cameras is n, where n is an integer greater than 1, for n RTSP video streams, the edge computing smart box deploys a decoding thread for each RTSP video stream, which is responsible for the hard decoding processing of the RTSP video stream, and deploys a corresponding algorithm reasoning thread for each RTSP video stream, which is responsible for completing the detection and analysis of perimeter intrusion targets based on the corresponding RTSP video stream, and deploys an alarm push thread for each RTSP video stream, which is responsible for sending the detection results in the form of a JSON file through an HTTP request to the sound and light alarm and the monitoring software platform.
[0009] Optionally, the detection results generated by the edge computing smart box include: target category, detection accuracy, detection box coordinate points, BASE64 image encoding, image size, IP camera ID, and timestamp.
[0010] Optionally, the detection model deployed in the edge computing smart box is an improved YOLOv8 neural network detection model. The improved YOLOv8 neural network detection model introduces an SE attention mechanism module in the channel dimension. The SE attention mechanism module improves the detection accuracy of small targets and complex scenes by enhancing the expression ability of key features; the SE attention mechanism module is introduced through the following steps: based on version 8.1.0 of YOLOv8, add the defined SEAttention(nn.Module) function in the nn / modules / conv.py file; import the SEAttention class in the nn / _init_.py file; add the judgment statement elif min{SEAttention} in the nn / task.py file to realize dynamic checking of the SEAttention model module, and enable specific logic to construct the network layer, thereby introducing the SE attention mechanism module.
[0011] Optionally, the improved YOLOv8 neural network detection model is trained and deployed through the following steps: construct a sample data set based on the YOLOv8n.pt model and data sets collected from the perimeters of different areas, then label the sample data in the sample data set through Labelimg software, label the category as person, and store the sample data in the format of the YOLO data set; shuffle the order of the sample data in the sample data set through the random.shuffle() function in Python, and divide it into a training set and a validation set according to a preset ratio; based on the YOLOv8n.pt model, first train on the person category of the COCO data set to obtain the coco_person.pt model, then use the coco_person.pt model as a pre-trained model, and further train on the constructed training set and validation set to obtain the best detection weight file best.pt; on an Ubuntu with x86_64 architecture On the 20.01 operating system, create a Docker container, which integrates the operating environment required for model compilation and quantization. Use the Docker container to compile and quantize the best detection weight file best.pt, and convert it into a model file that can be run on the edge computing smart box through the MLIR tool; migrate the model file to the edge computing smart box, build the program running environment according to the multi-threaded mode, and write the running script start.sh for program startup and monitoring.
[0012] Optionally, the best detection weight file best.pt is compiled and quantized using a Docker container, and converted into a model file that can be run on an edge computing smart box through an MLIR tool, including: calling the YOLO class in the UltralyticsYOLOv8 framework to load the best detection weight file best.pt; using the export method to export the model file best.onnx in a dynamic ONNX format, and then using the MLIR tool to convert it into a quantized inference model file that can be run on an edge computing smart box.
[0013] Optionally, when running the start.sh script for program startup and monitoring, execute the command of the start.sh script based on the Linux ARM64 architecture, enter the specified directory, start the program startup process and the model service daemon process. If any process is not running, the start.sh script will restart the unstarted process after a preset time. Use the ps command to check the process status and nohup to start the process to ensure system stability and continuous service.
[0014] Optionally, different areas include residential areas, industrial areas, commercial areas, mountainous areas, and reservoir areas, and sample data in data sets collected from perimeters of different areas include daytime sample data, nighttime sample data, sample data of different weather conditions, and sample data of different seasons.
[0015] Optionally, when the coco_person.pt model is used as a pre-trained model and further trained on the constructed training set and validation set to obtain the best detection weight file best.pt, the Adam optimizer is selected, the initial learning rate is set to 0.003, the batch size is set to 24, the maximum number of iterations is set to 800, and the maximum training market is set to 30 hours.
[0016] The embodiment of the present application proposes a perimeter intrusion target monitoring system, which installs multiple IP cameras on fixed poles arranged at preset distances around the perimeter of the target area to achieve full coverage of the viewing angle of the perimeter of the target area, and perimeter intrusion targets in any direction can be detected. An improved YOLOv8 neural network detection model is deployed in the edge computing smart box to detect and analyze perimeter intrusion targets. The SE attention mechanism is introduced in the improved YOLOv8 neural network detection model, which significantly improves the detection accuracy of small targets and complex scenes. The edge computing smart box can perform parallel detection and analysis of multi-channel video data, effectively improving the real-time recognition capability of perimeter intrusion targets. Through edge computing processing, it reduces the dependence on cloud servers and network bandwidth, avoids delays and data loss caused by bandwidth limitations, especially in places with poor network environments such as mountainous areas and reservoirs, ensuring the stability and efficiency of the system. Combined with the sound and light alarm and the monitoring software platform, it can trigger an alarm and display the alarm screen in time when a perimeter intrusion target is found, greatly improving the response speed and protection capability of perimeter security. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] One or more embodiments are exemplarily described by the pictures in the corresponding drawings, and these exemplary descriptions do not constitute limitations on the embodiments.
[0018] Figure 1 is a structural schematic diagram of a perimeter intrusion target monitoring system provided in one embodiment of the present application;
[0019] Figure 2 is a schematic diagram of an implementation of a perimeter intrusion target monitoring system provided in one embodiment of the present application;
[0020] Figure 3 It is a schematic diagram of the detection effect provided in one embodiment of the present application. DETAILED DESCRIPTION
[0021] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in detail below in conjunction with the accompanying drawings. However, it will be appreciated by those skilled in the art that in the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, even without these technical details and various changes and modifications based on the following embodiments, the technical scheme claimed in the present application can also be implemented. The division of the following embodiments is for the convenience of description, and the specific implementation of the present application should not be construed as any limitation, and the various embodiments can be combined and referenced with each other under the premise of no contradiction.
[0022] An embodiment of the present application proposes a perimeter intrusion target monitoring system. The implementation details of the perimeter intrusion target monitoring system proposed in this embodiment are specifically described below. The following content is only the implementation details provided for the convenience of understanding and is not necessary for the implementation of this solution.
[0023] The specific structure of the perimeter intrusion target monitoring system proposed in this embodiment can be as follows: Figure 1 As shown, including: multiple IP cameras ( Figure 1 Four IP cameras are shown, namely IP camera 11, IP camera 12, IP camera 13 and IP camera 14), power supply equipment 20, switch 30, edge computing smart box 40, sound and light alarm 50 and monitoring software platform 60. Multiple IP cameras are installed on fixed poles arranged at preset distances around the perimeter of the target area to achieve full coverage of the viewing angle of the perimeter of the target area.
[0024] Each IP camera is used to collect video data in real time. The power supply device 20 is used to provide power support for all IP cameras. A communication connection is established between the switch 30 and each IP camera, which is used to transmit the video data collected by each IP camera to the edge computing smart box 40 and the monitoring software platform 60 through a self-organizing network. The edge computing smart box 40 is used to detect and analyze multi-channel video data, and send the detection results and alarm information to the sound and light alarm 50 and the monitoring software platform 60. The sound and light alarm 50 is used to trigger the sound and light alarm after receiving the detection results and alarm information, and emit strong light and sound waves to drive away the perimeter intrusion target. The monitoring software platform 60 is used to display the real-time monitoring screen corresponding to each IP camera, and after receiving the detection results and alarm information, it focuses on displaying the alarm content and the real-time monitoring screen corresponding to the IP camera that triggered the alarm, so that the on-duty personnel can understand the perimeter defense situation in a timely manner.
[0025] In one example, the spacing between the fixed poles is 50 meters.
[0026] Figure 2 This is a schematic diagram of a perimeter intrusion target monitoring system proposed in this embodiment. Figure 2 A detailed introduction is given to the various components of the border intrusion target monitoring system.
[0027] In the perimeter intrusion target monitoring system, each IP camera transmits the real-time collected video data to the switch in the form of RTSP video stream according to the real-time streaming protocol, and then transmits it to the edge computing smart box and monitoring software platform. The edge computing smart box sends the detection results and alarm information in the form of JSON files through HTTP requests to the sound and light alarm and monitoring software platform. The alarm information received by the sound and light alarm and monitoring software platform is HTTP alarm information.
[0028] The edge computing smart box is deployed with pre-trained detection models, such as Figure 2 As shown in the figure, the edge computing smart box uses a multi-threaded mode to implement the full-process algorithm reasoning for multiple RTSP video streams. Assume that the number of IP cameras is n, where n is an integer greater than 1 (if the video stream does not extract frames, then n≤8, if the video stream extracts frames, then n≤16). For n RTSP video streams, the edge computing smart box deploys a decoding thread (Thread-Decode) for each RTSP video stream, which is responsible for the hard decoding processing of the RTSP video stream. Correspondingly, the edge computing smart box deploys a corresponding algorithm reasoning thread (Thread-Det) for each RTSP video stream, which is responsible for completing the detection and analysis of perimeter intrusion targets based on the corresponding RTSP video stream. At the same time, the edge computing smart box also deploys an alarm push thread (Thread-push) for each RTSP video stream, which is responsible for transmitting the detection results in the form of a JSON file through an HTTP request to the sound and light alarm and the monitoring software platform.
[0029] In the perimeter intrusion target monitoring system, the detection results generated by the edge computing smart box include but are not limited to target category, detection accuracy, detection box coordinate points, BASE64 image encoding, image size, IP camera ID, and timestamp.
[0030] In one example, the detection model deployed in the edge computing smart box is an improved YOLOv8 neural network detection model. The improved YOLOv8 neural network detection model introduces the SE attention mechanism module in the channel dimension. The SE attention mechanism module improves the detection accuracy of small targets and complex scenes by enhancing the expression ability of key features. The SE attention mechanism module is introduced through the following steps: First, based on the 8.1.0 version of YOLOv8, add the defined SEAttention(nn.Module) function in the nn / modules / conv.py file; then, import the SEAttention class in the nn / _init_.py file; finally, add the judgment statement elif min{SEAttention} in the nn / task.py file to dynamically check the SEAttention model module, and enable specific logic to construct the network layer, thereby introducing the SE attention mechanism module.
[0031] In one example, the improved YOLOv8 neural network detection model needs to be trained before the perimeter intrusion target monitoring system is built.
[0032] The first thing to do is to build a training set and a validation set. A sample data set is built based on the YOLOv8n.pt model and data sets collected from the perimeters of different regions. Then, the sample data in the sample data set is labeled with the label category person through the Labelimg software, and the sample data is stored in the format of the YOLO data set. Next, the order of the sample data in the sample data set is disrupted through the random.shuffle() function in Python, and the training set and validation set are obtained according to the preset ratio. Generally speaking, the preset ratio is 8:2. Assuming that a total of 31,500 sample data are collected, 25,200 of them constitute the training set, and the remaining 6,300 constitute the validation set.
[0033] It is worth noting that different areas include but are not limited to residential areas, industrial areas, commercial areas, mountainous areas, and reservoir areas. The sample data in the data sets collected from the perimeters of different areas include daytime sample data, nighttime sample data, sample data in different weather conditions, sample data in different seasons, etc.
[0034] The next step is model training. Based on the YOLOv8n.pt model, basic training is first performed on the person category of the COCO dataset to obtain the coco_person.pt model. Then, the coco_person.pt model is used as a pre-trained model and further trained on the constructed training set and validation set to obtain the best detection weight file best.pt. In further training, the Adam optimizer is selected, the initial learning rate is set to 0.003, the batch size is set to 24, the maximum number of iterations is set to 800, and the maximum training period is set to 30 hours.
[0035] Next, we need to generate, compile, and deploy the model file. On the Ubuntu 20.01 operating system of the x86_64 architecture, create a Docker container that integrates the operating environment required for model compilation and quantization. Use the Docker container to compile and quantize the best detection weight file best.pt, and use the MLIR tool to convert it into a model file that can be run on the edge computing smart box. Finally, migrate the model file to the edge computing smart box, build the program running environment based on the multi-threaded mode, and write the running script start.sh for program startup and monitoring.
[0036] In an example, when generating a model file, you first need to call the YOLO class in the Ultralytics YOLOv8 framework, load the best detection weight file best.pt, then use the export method to export the model file best.onnx in the dynamic ONNX format, and finally use the MLIR tool to convert it into a model file for quantized reasoning that can be run on the edge computing smart box.
[0037] In one example, when running the start.sh script for program startup and monitoring, you need to execute the command of the start.sh script based on the LinuxARM64 architecture to enter the specified directory, start the program startup process and the model service daemon process. If any process is not running, the start.sh script will restart the unstarted process after a preset time (usually set to 5 seconds). After the process is started, you need to use the ps command to check the process status and nohup to start the process to ensure system stability and continuous service.
[0038] In one example, the detection effect of perimeter intrusion targets can be as follows: Figure 3 As shown, perimeter intrusion targets can be detected promptly and accurately in various environments, weather conditions and time periods.
[0039] A perimeter intrusion target monitoring system proposed in this embodiment installs multiple IP cameras on fixed poles arranged at preset distances around the perimeter of the target area to achieve full coverage of the viewing angle of the perimeter of the target area, and perimeter intrusion targets in any direction can be detected. An improved YOLOv8 neural network detection model is deployed in the edge computing smart box to detect and analyze perimeter intrusion targets. The SE attention mechanism is introduced in the improved YOLOv8 neural network detection model, which significantly improves the detection accuracy of small targets and complex scenes. The edge computing smart box can perform parallel detection and analysis of multi-channel video data, effectively improving the real-time recognition capability of perimeter intrusion targets. Through edge computing processing, it reduces the dependence on cloud servers and network bandwidth, avoids delays and data loss caused by bandwidth limitations, especially in places with poor network environments such as mountainous areas and reservoirs, ensuring the stability and efficiency of the system. Combined with the sound and light alarm and the monitoring software platform, it can trigger an alarm and display the alarm screen in time when a perimeter intrusion target is found, greatly improving the response speed and protection capability of perimeter security.
[0040] It is worth mentioning that all modules involved in this embodiment are logic modules. In practical applications, a logic unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed by this application, but this does not mean that there are no other units in this embodiment.
[0041] Those skilled in the art will appreciate that the above embodiments are specific embodiments for implementing the present application, and in actual applications, various changes may be made thereto in form and detail without departing from the spirit and scope of the present application.
Claims
1. A perimeter intrusion target monitoring system, characterized in that: include: Multiple IP cameras, power supply equipment, switches, edge computing smart boxes, sound and light alarms, and monitoring software platforms. Multiple IP cameras are installed on fixed poles at preset intervals around the perimeter of the target area to achieve full coverage of the perimeter of the target area; IP cameras are used to collect video data in real time; The power supply equipment is used to provide power support for all IP cameras; A communication connection is established between the switch and each IP camera, which is used to transmit the video data collected by each IP camera to the edge computing smart box and monitoring software platform through an ad hoc network; The edge computing smart box is used to detect and analyze multi-channel video data, and send the detection results and alarm information to the sound and light alarm and monitoring software platform; The sound and light alarm is used to trigger the sound and light alarm after receiving the detection results and alarm information, and emit strong light and sound waves to drive away the perimeter intrusion targets; The monitoring software platform is used to display the real-time monitoring screen corresponding to each IP camera, and after receiving the detection results and alarm information, it focuses on displaying the alarm content and the real-time monitoring screen corresponding to the IP camera that triggered the alarm.
2. A perimeter intrusion target monitoring system according to claim 1, characterized in that: Each IP camera transmits the real-time collected video data to the switch in the form of RTSP video stream according to the real-time streaming protocol, and then transmits it to the edge computing smart box and monitoring software platform; The edge computing smart box sends the detection results and alarm information in the form of JSON files through HTTP requests to the sound and light alarm and monitoring software platform.
3. A perimeter intrusion target monitoring system according to claim 2, characterized in that: A pre-trained detection model is deployed in the edge computing smart box, which uses a multi-threaded mode to implement full-process algorithm reasoning for multiple RTSP video streams; Assume that the number of IP cameras is n, where n is an integer greater than 1. For n RTSP video streams, the edge computing smart box deploys a decoding thread for each RTSP video stream, which is responsible for the hard decoding processing of the RTSP video stream. A corresponding algorithm reasoning thread is deployed for each RTSP video stream, which is responsible for completing the detection and analysis of perimeter intrusion targets based on the corresponding RTSP video stream. An alarm push thread is deployed for each RTSP video stream, which is responsible for sending the detection results in the form of a JSON file through an HTTP request to the sound and light alarm and the monitoring software platform.
4. A perimeter intrusion target monitoring system according to claim 3, characterized in that: The detection results generated by the edge computing smart box include: target category, detection accuracy, detection box coordinate points, BASE64 image encoding, image size, IP camera ID, and timestamp.
5. A perimeter intrusion target monitoring system according to claim 3, characterized in that: The detection model deployed in the edge computing smart box is an improved YOLOv8 neural network detection model. The improved YOLOv8 neural network detection model introduces the SE attention mechanism module in the channel dimension. The SE attention mechanism module improves the detection accuracy of small targets and complex scenes by enhancing the expression ability of key features. The SE attention mechanism module is introduced through the following steps: Based on version 8.1.0 of YOLOv8, add the defined SEAttention(nn.Module) function in the nn / modules / conv.py file; Import the SEAttention class in the nn / _init_.py file; Add the judgment statement elif min{SEAttention} in the nn / task.py file to dynamically check the SEAttention model module and enable specific logic to construct the network layer, thereby introducing the SE attention mechanism module.
6. A perimeter intrusion target monitoring system according to claim 3, characterized in that: The improved YOLOv8 neural network detection model is trained and deployed through the following steps: A sample dataset is constructed based on the YOLOv8n.pt model and datasets collected from different area perimeters. The sample data in the sample dataset is then labeled with the label category person using Labelimg software, and the sample data is centrally stored in the format of the YOLO dataset. The random.shuffle() function in Python is used to shuffle the order of the sample data in the sample data set, and the training set and the validation set are divided according to the preset ratio. Based on the YOLOv8n.pt model, first train it on the person category of the COCO dataset to obtain the coco_person.pt model. Then use the coco_person.pt model as a pre-trained model and further train it on the constructed training set and validation set to obtain the best detection weight file best.pt. On the Ubuntu 20.01 operating system of the x86_64 architecture, create a Docker container that integrates the operating environment required for model compilation and quantization. Use the Docker container to compile and quantize the best detection weight file best.pt, and use the MLIR tool to convert it into a model file that can be run on the edge computing smart box. Migrate the model file to the edge computing smart box, build the program running environment based on the multi-threaded mode, and write the start.sh running script for program startup and monitoring.
7. A perimeter intrusion target monitoring system according to claim 6, characterized in that: Use the Docker container to compile and quantize the best detection weight file best.pt, and convert it into a model file that can be run on the edge computing smart box through the MLIR tool, including: Call the YOLO class in the Ultralytics YOLOv8 framework and load the best detection weight file best.pt; Use the export method to export the model file best.onnx in the dynamic ONNX format, and then use the MLIR tool to convert it into a model file for quantized reasoning that can be run on the edge computing smart box.
8. A perimeter intrusion target monitoring system according to claim 6, characterized in that: When running the start.sh script for program startup and monitoring, execute the command of the start.sh script based on the Linux ARM64 architecture, enter the specified directory, start the program startup process and the model service daemon process. If any process is not running, the start.sh script will restart the unstarted process after the preset time. Ensure system stability and continuous service by using the ps command to check process status and nohup to start the process.
9. A perimeter intrusion target monitoring system according to claim 6, characterized in that: Different areas include residential areas, industrial areas, commercial areas, mountainous areas, and reservoir areas. The sample data in the data set collected from the perimeters of different areas include daytime sample data, nighttime sample data, sample data in different weather conditions, and sample data in different seasons.
10. A perimeter intrusion target monitoring system according to any one of claims 6 to 9, characterized in that: When using the coco_person.pt model as the pre-trained model and further training on the constructed training set and validation set to obtain the best detection weight file best.pt, the Adam optimizer was selected, the initial learning rate was set to 0.003, the batch size was set to 24, the maximum number of iterations was set to 800, and the maximum training period was set to 30 hours.