Electric power operation safety area detection method based on infrared image
By combining infrared image enhancement and an improved YOLOv8 detection model with the SORT tracking algorithm, the stability problem of visible light image detection at night and in complex environments has been solved, enabling all-weather, all-time intrusion detection and improving recognition accuracy and system stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-17
AI Technical Summary
Existing technologies for intrusion detection systems based on visible light images suffer from poor stability and reliability at night or in complex environments, making it difficult to achieve continuous perception and accurate identification of intrusion behavior around the clock.
By combining infrared image enhancement technology and an improved YOLOv8 target detection model with the SORT multi-target tracking algorithm, images are acquired through an infrared camera, preprocessed and feature-enhanced, channel attention and spatial attention modules are used to improve target recognition accuracy, and temporal correlation is performed through the SORT algorithm to achieve stable intrusion detection.
It achieves continuous sensing capabilities around the clock, even at night or in complex environments, improving the accuracy and stability of intrusion detection, reducing computational complexity, and making it suitable for low-power embedded platforms.
Smart Images
Figure CN121686513A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power operation safety monitoring technology, and in particular to a method for detecting safe areas in power operations based on infrared images. Background Technology
[0002] With rapid societal development and the continuous improvement of industrial intelligence, the demand for security protection, production monitoring, and intelligent security in public places is increasing daily. To ensure the safety of personnel and equipment, area intrusion detection technology is widely used in factory work areas, hazardous locations, warehousing and logistics, transportation hubs, and military protected zones. This technology aims to perceive dynamic changes within a monitored area in real time through image sensors or video surveillance systems, detect whether unauthorized targets have entered a preset warning range, and promptly issue an alarm upon identifying intrusion behavior to prevent security incidents or illegal intrusion events.
[0003] The key to area intrusion detection lies in the automatic detection, tracking, and behavior recognition of targets within the monitored scene. Traditional intrusion detection systems mostly rely on target detection techniques based on visible light images, using background modeling, frame differencing, or deep learning methods to identify targets entering the protected area. However, visible light images are highly susceptible to external lighting conditions. In environments such as nighttime, backlighting, rain, snow, or fog, target contrast is significantly reduced, edge features are blurred, and detection algorithms are prone to missed detections, false detections, or even complete failure. Furthermore, low-light environments can lead to increased noise and shadow interference, significantly reducing system stability and reliability.
[0004] Infrared thermal imaging (IR) technology, as a sensing method based on thermal radiation signals, can achieve all-weather imaging without relying on ambient light sources. It detects the infrared energy distribution on a target surface, converting thermal differences into grayscale or pseudo-color images, effectively identifying targets such as people and animals even in complete darkness or complex weather conditions. Compared to traditional visible light surveillance, infrared imaging systems offer advantages such as strong anti-interference capabilities, low-light adaptability, and stable reliability, making them an ideal approach for intelligent monitoring in nighttime and complex scenarios. Therefore, intrusion detection methods based on infrared images have become an important development direction in the field of intelligent security.
[0005] In recent years, the rapid development of deep learning has provided strong technical support for target detection and behavior recognition tasks. Convolutional neural network (CNN) structures, represented by models such as the YOLO series, SSD, and Faster R-CNN, have been applied in the field of image detection. In intrusion detection scenarios, detection models based on the YOLO series are commonly used to achieve real-time identification and tracking. The journal *China Safety Science and Technology*, 2024, 20(07), proposed a "Dynamic Dangerous Area Intrusion Detection Method Based on Deep Learning." This method detects intrusions in dynamic dangerous areas based on an improved YOLOv5 model. It uses binocular vision technology to acquire depth information, delineates dynamic dangerous areas through depth threshold filtering, K-means clustering, and connected component processing, and employs an improved YOLOv5 model that integrates a lightweight PDConv convolutional module and a convolutional attention mechanism at the detection end to detect and identify targets. Finally, the intrusion status is determined based on the IOU of the target bounding box. This method significantly reduces computational complexity and the number of parameters while ensuring detection accuracy, and improves the model's deployment capability on industrial edge devices. However, this method is essentially a static frame-level judgment mechanism, lacking information correlation and state persistence over time. It heavily relies on the detection model's performance in a single frame; when there is brief target occlusion, missed detection, or fluctuations in inter-frame recognition, the system will be unable to correctly identify persistent intrusion behavior. Furthermore, it still depends on visible light cameras, and its detection performance remains unstable at night or in adverse environments due to limitations in lighting conditions and occlusion.
[0006] Another approach proposes an integrated intrusion detection system that combines detection, tracking, and re-identification. This system achieves precise control based on a collaborative workflow of YOLOv7 target detection, target tracking, and pedestrian re-identification. The process begins with frame-by-frame analysis of the video stream using a target detection model to locate and select all personnel and vehicles in the frame in real time. Subsequently, the system introduces the DeepSORT multi-target tracking algorithm, assigning a unique ID to each detected target and continuously tracking its trajectory, thus connecting isolated detection points into a stable trajectory for precise control of illegal targets. Finally, an alarm is triggered immediately when the trajectory of any identified illegal identity or target enters the warning area. Its advantage lies in compensating for occasional errors between detection frames through temporal correlation, making intrusion judgment more stable. While this approach offers superior performance in multi-target tracking and identity recognition, its drawbacks include high deployment costs and strict requirements for scenario adaptability. To ensure tracking accuracy, this approach employs a deep network structure and a pedestrian re-identification module, resulting in a large number of model parameters and high computational complexity, making it difficult to implement in real-time on low-power embedded platforms. Secondly, this solution relies entirely on RGB video streams, and its sensing capabilities will significantly decrease or even fail in environments with drastic changes in lighting, low visibility, or complete darkness at night. This fundamental weakness limits its application in industrial security scenarios that require stable operation around the clock. Summary of the Invention
[0007] The purpose of this invention is to propose a method for detecting safe areas in power operations based on infrared images that can overcome the limitations of visible light perception while ensuring continuous perception capabilities in different scenarios.
[0008] The technical solution of this invention is as follows: A region intrusion detection method based on infrared images is implemented through the following steps: Step 1: Publish the images captured by the infrared camera through the ROS topic and subscribe to the real-time infrared images input from the infrared camera sensor; Step 2: Preprocess and enhance the input raw infrared image; Step 3: Based on infrared image data, construct an improved YOLOv8 target detection model; Step 4: Input the enhanced infrared image into the improved YOLOv8 model for detection. The model output includes pedestrian categories and their two-dimensional bounding box coordinates. Step 5: Apply the SORT algorithm to the detection results obtained in Step 4 for multi-target tracking; Step 6: Extract the coordinates of the bottom center point of the bounding box of the target detected in each frame, and determine whether the point falls into the set polygon area; if the detection results of multiple consecutive frames meet the entry conditions, it is determined that an intrusion has occurred. Step 7: Once the intrusion event is confirmed as genuine, the system immediately generates an alarm signal and publishes it to the ROS message channel.
[0009] Furthermore, in step two, the infrared image obtained in step one is subjected to Gaussian filtering to remove noise interference from the image; then, the image is divided into 8×8 regions, and the average gray value and brightness range of each region are calculated. The enhancement coefficient is dynamically adjusted according to the overall brightness state of the region to achieve local brightness balance; then, the brightness range of each region is linearly or piecewise mapped to redistribute the gray values to the standard display range, and the overall brightness level is adjusted by Gamma correction. Finally, the adaptive histogram equalization (CLAHE) algorithm is used to locally enhance the image and improve edge and texture details.
[0010] Furthermore, in step three, a channel attention module (CAM module) and a spatial attention module (SAM module) are introduced into the backbone network of the YOLOv8 object detection model. In the backbone feature extraction stage, the C2fGhost module is used to extract features instead of the original C2f structure, so as to reduce the amount of computation and maintain the feature representation ability.
[0011] Furthermore, the channel attention module is placed in the first layer of the backbone network, the C2fGhost structure, to complete the initial extraction of edge and heat intensity information, and the output features are: The channel attention module focuses on this feature. Global average pooling and global maximum pooling are performed for spatial dimensionality reduction. The features from both pooling methods are then input into a multilayer perceptron (MLP) with shared weights to extract the nonlinear relationships between channels, thereby learning the importance of each channel and generating a channel weight matrix. Finally, the weight matrix and original features Multiply to obtain the weighted input features. :
[0012]
[0013] in, The activation function is Sigmoid; the MLP consists of two fully connected layers. For global average pooling; For global maximum pooling; Element-wise product; The features are channel attention-weighted; the spatial attention module is embedded after the second C2fGhost module in the backbone network, located in the high-level semantic feature extraction stage; the input feature map in this stage is... It already contains rich semantic information; the SAM module guides the model to focus on the target region by calculating the correlation in the spatial dimension, and its calculation formula is as follows:
[0014]
[0015] in, Indicates channel splicing; Indicates the convolution operation; Use the Sigmoid activation function; For global average pooling; For global maximum pooling; Represents element-wise product; This is the feature map after spatial attention weighting.
[0016] This invention, based on the YOLOv8 detection framework, designs a detection network structure that integrates the CAM (Channel Attention) and SAM (Spatial Attention) mechanisms. The CAM module models inter-channel dependencies using lightweight one-dimensional convolution, enabling the network to automatically focus on key thermal signal channels. The SAM module enhances the spatial response of the target region and suppresses background noise through spatial grouping enhancement. The combination of these two mechanisms allows the model to generate significant feature responses even under weak infrared texture conditions, improving the accuracy of blurred target recognition.
[0017] Furthermore, in step five, the SORT algorithm process is as follows: the pedestrian position and bounding box in each frame detection result obtained in step four are input into the Kalman filter to model the motion state of the target and predict its position; then, the Hungarian matching algorithm is used to associate the detection result of the current frame with the predicted trajectory, assign a unique identifier to the continuously detected target, and compensate for short-term occlusion or missed detection through the state update mechanism.
[0018] Compared with existing technologies, this invention constructs a method and system for regional intrusion detection suitable for infrared imaging environments. Employing infrared imaging as the primary sensing means, it overcomes the problem of decreased perception capability of visible light surveillance in complex environments such as nighttime, backlight, rain, snow, and fog, achieving continuous perception capability in all weather and at all times. This invention uses image enhancement, an improved YOLOv8 detection network, and SORT multi-target tracking as its core, integrating target recognition, trajectory association, and region determination into a continuous and tightly integrated processing flow. Enhanced infrared images allow the model to more easily capture the true contours of pedestrians, while the improved YOLOv8 exhibits better recognition stability for blurred, scale-varying, or low-contrast targets in infrared scenes. The detection results are then temporally associated by SORT, ensuring that each target has a reliable trajectory in consecutive frames, preventing immediate loss even in the event of occasional brief occlusion. Finally, the system makes a judgment based on the relationship between the bottom center point of the target bounding box and the designated area during tracking, thus balancing the requirements of recognition accuracy and false alarm suppression, making intrusion determination more reliable and natural. Attached Figure Description
[0019] Figure 1 This is a flowchart illustrating the electric work safety zone detection method based on infrared images of the present invention. Figure 2 This is a flowchart of the infrared image preprocessing process of the present invention; Figure 3 This is a diagram of the improved Yolov8 network structure of this invention; Figure 4 This is a schematic diagram of the structure of the electric work safety area detection system based on infrared images of the present invention; Figure 5 This is a diagram illustrating intrusion detection by personnel. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] A region intrusion detection method based on infrared images, such as Figure 1 As shown, this can be achieved through the following steps: Step 1: Publish the images captured by the infrared camera through the ROS topic and subscribe to the real-time infrared images input from the infrared camera sensor; Step 2: Preprocess and enhance the input raw infrared image; Step 3: Based on infrared image data, construct an improved YOLOv8 target detection model; Step 4: Input the enhanced infrared image into the improved YOLOv8 model for detection. The model output includes pedestrian categories and their two-dimensional bounding box coordinates. Step 5: Apply the SORT algorithm to the detection results obtained in Step 4 for multi-target tracking; Step 6: Extract the coordinates of the bottom center point of the bounding box of the target detected in each frame, and determine whether the point falls into the set polygon area; if the detection results of multiple consecutive frames meet the entry conditions, it is determined that an intrusion has occurred. Step 7: Once the intrusion event is confirmed as genuine, the system immediately generates an alarm signal and publishes it to the ROS message channel.
[0022] In one embodiment of the present invention, such as Figure 2 As shown, in step two, the infrared image obtained in step one is subjected to Gaussian filtering to remove noise interference from the image; then, the image is divided into 8×8 regions, and the average gray value and brightness range of each region are calculated. The enhancement coefficient is dynamically adjusted according to the overall brightness state of the region to achieve local brightness balance; then, the brightness range of each region is linearly or piecewise mapped to redistribute the gray values to the standard display range, and the overall brightness level is adjusted by Gamma correction. Finally, the adaptive histogram equalization (CLAHE) algorithm is used to locally enhance the image and improve edge and texture details.
[0023] In one embodiment of the present invention, such as Figure 3 As shown, in step three, a channel attention module (CAM module) and a spatial attention module (SAM module) are introduced into the backbone network of the YOLOv8 object detection model. In the backbone feature extraction stage, the C2fGhost module is used to extract features instead of the original C2f structure, so as to reduce the amount of computation and maintain the feature expression ability. The channel attention module is placed in the first layer of the backbone network, the C2fGhost structure, to perform preliminary edge and heat intensity information extraction, and the output features are: The channel attention module focuses on this feature. Global average pooling and global maximum pooling are performed for spatial dimensionality reduction. The features from both pooling methods are then input into a multilayer perceptron (MLP) with shared weights to extract the nonlinear relationships between channels, thereby learning the importance of each channel and generating a channel weight matrix. Finally, the weight matrix and original features Multiply to obtain the weighted input features. :
[0024]
[0025] in, The activation function is Sigmoid; the MLP consists of two fully connected layers. For global average pooling; For global maximum pooling; Element-wise product; The features are channel attention-weighted; the spatial attention module is embedded after the second C2fGhost module in the backbone network, located in the high-level semantic feature extraction stage; the input feature map in this stage is... It already contains rich semantic information; the SAM module guides the model to focus on the target region by calculating the correlation in the spatial dimension, and its calculation formula is as follows:
[0026]
[0027] in, Indicates channel splicing; Indicates the convolution operation; Use the Sigmoid activation function; For global average pooling; For global maximum pooling; Represents element-wise product; This is the feature map after spatial attention weighting.
[0028] In one embodiment of the present invention, in step five, the SORT algorithm process is as follows: the pedestrian position and bounding box in each frame detection result obtained in step four are input into the Kalman filter to model the motion state of the target and predict its position; then the Hungarian matching algorithm is used to associate the detection result of the current frame with the predicted trajectory, assign a unique identifier to the continuously detected target, and compensate for short-term occlusion or missed detection through a state update mechanism.
[0029] like Figure 4 As shown, a power operation safety zone detection system based on infrared images includes: The acquisition module publishes images captured by the infrared camera through ROS topics and subscribes to real-time infrared images input from the infrared camera sensor. The image enhancement module preprocesses and enhances the input raw infrared image; The YOLOv8 detection module constructs an improved YOLOv8 target detection model based on infrared image data. The enhanced infrared image is input into the improved YOLOv8 model for detection, and the model output includes pedestrian categories and their two-dimensional bounding box coordinates. The SORT tracking module applies the SORT algorithm to the detection results of the YOLOv8 detection module for multi-target tracking. The intrusion detection module extracts the coordinates of the bottom center point of the bounding box of the detected target in each frame and determines whether the point falls into the set polygon area. If the detection results of multiple consecutive frames meet the entry conditions, an intrusion is determined to have occurred. When the intrusion event is determined to be true, the system immediately generates an alarm signal and publishes it to the ROS message channel.
[0030] like Figure 5 As shown, after receiving a frame of raw infrared image through the acquisition module, the system first performs noise reduction and contrast optimization by the image enhancement module to make the pedestrian thermal imaging outline clearer and more stable. The enhanced image is then sent to the YOLOv8 detection module, which quickly identifies the "person" target and outputs the bounding box coordinates (center point coordinates (320, 180), width 80 pixels, height 140 pixels). The detection result enters the SORT tracking module, which performs smoothing through Kalman filtering to further eliminate the slight flicker and frame jitter common in infrared images. The smoothed target center point (320, 180) is sent to the intrusion detection module, which performs intersection judgment with the preset yellow warning polygon area. If the target is found to have entered the restricted area, the system immediately draws a red highlighted intrusion box and a yellow warning area on the output image, and overlays a conspicuous "INTRUSION!" in the upper right corner. At the same time, an alarm signal is issued through the ROS topic, completing an efficient and stable nighttime intrusion detection process.
Claims
1. A method for area intrusion detection based on infrared images, characterized in that, The implementation is achieved by the following steps: Step one, publish the pictures taken by the infrared camera through the ROS topic, and subscribe to the real-time infrared image input by the infrared camera sensor; Step two, preprocess and enhance the input raw infrared image; Step three, based on the infrared image data, construct an improved YOLOv8 target detection model; Step four, input the enhanced infrared image into the improved YOLOv8 model for detection, and the model output contains the category of pedestrians and their two-dimensional bounding box coordinates; Step five, apply the SORT algorithm to the detection results obtained in step four for multi-target tracking; Step six, extract the center point coordinates of the bottom of the bounding box in each frame of the detected target, and judge whether the point falls into the set polygon region; If the continuous multi-frame detection results meet the entering condition, it is determined that an intrusion behavior occurs; Step seven, if the intrusion event is determined to be true, the system immediately generates an alarm signal and publishes it to the ROS message channel.
2. The method of claim 1, wherein, In step two, Gaussian filtering is performed on the infrared image obtained in step one to remove noise interference in the image; Then, the image is divided into 8x8 regions, and the average gray value and brightness range of each region are calculated, and the enhancement coefficient is dynamically adjusted according to the overall brightness state of the region to achieve local brightness equalization; then the linear or segmented mapping of the brightness range of each region is performed, and the gray value is redistributed to the standard display interval, and the overall brightness level is adjusted through Gamma correction, and finally the adaptive histogram equalization (CLAHE) algorithm is used for local enhancement of the image to improve the edge and texture details.
3. The method of claim 1, wherein, In step three, the channel attention module (CAM module) and the spatial attention module (SAM module) are introduced into the backbone network of the YOLOv8 target detection model, and in the backbone feature extraction stage, the C2fGhost module is used for extraction instead of the original C2f structure to reduce the calculation amount and maintain the feature expression ability.
4. The method of claim 3, wherein, The channel attention module is placed in the first layer C2fGhost structure of the backbone network, completes preliminary edge and heat intensity information extraction, and outputs features ; The channel attention module performs global average pooling and global maximum pooling on the features , performs spatial dimension reduction processing, inputs two kinds of pooled features into a multi-layer perceptron (MLP) shared weight, extracts inter-channel nonlinear relationships, learns the importance of each channel, and generates a channel weight matrix ; Finally, the weight matrix is multiplied by the original features to obtain weighted input features : wherein, is a Sigmoid activation function; the MLP consists of two fully connected layers; is global average pooling; is global maximum pooling; is an element-wise product; is a channel attention weighted feature; the spatial attention module is embedded after the second C2fGhost module of the backbone network, located in the high-level semantic feature extraction stage; the input feature map at this stage contains rich semantic information; the SAM module guides the model to focus on the target area by calculating the correlation in the spatial dimension, and its calculation formula is as follows: wherein, represents channel concatenation; represents convolution operation; is a Sigmoid activation function; is global average pooling; is global maximum pooling; represents an element-wise product; is a spatial attention weighted feature map.
5. The method of claim 1, wherein In step five, the flow of the SORT algorithm is as follows: input the position and bounding box of the pedestrian in each frame of the detection result obtained in step four into the Kalman filter, model the motion state of the target and predict the position; Then use the Hungarian matching algorithm to associate the detection results of the current frame with the predicted trajectory, assign a unique identifier to the continuously detected target, and compensate for short-term occlusion or missed detection through the state update mechanism.