Double-camera personnel security detection method

Abstract

The invention belongs to the technical field of personnel security and protection detection, and discloses a double-camera personnel security and protection detection method, which comprises the following specific steps: step 1, deploying and installing an RGB camera and a thermal infrared camera on double cameras; through cooperative deployment and accurate calibration of the RGB camera and the thermal infrared camera, all-weather and all-scene reliable monitoring is realized, and a detection blind area of a single sensor in a complex environment is effectively overcome; the deep learning Retinanet target detection algorithm can greatly improve the personnel target detection rate in the image by using a low threshold value, and further improves the detection accuracy and robustness in combination with hot red imaging; the image is processed through the Gaussian mixture model, so that the stable performance can be kept regardless of illumination variation, temperature fluctuation or equipment interference, and excellent universality and expandability are shown; personnel safety protection is realized through alarm and shutdown control, and the accuracy of personnel detection in a dangerous area is improved on the whole.

Description

Technical Field

[0001] The present invention belongs to the technical field of personnel security detection, and specifically relates to a dual-camera personnel security detection method. Background Technique

[0002] In the modern industrial production environment, the safety management of various dangerous areas has always been the primary task of ensuring the safety of personnel's lives and the stable operation of equipment. These dangerous areas mainly include, but are not limited to: high-speed rotating mechanical operation areas, high-voltage electrical equipment areas, high-temperature and high-pressure operation areas, storage areas of toxic and harmful substances, and automated logistics transportation channels, etc. There are various high-risk factors such as mechanical injuries, electric shocks, burns, and object strikes in these areas. Once a person strays into or enters illegally, it is very likely to cause serious personal injury accidents and significant property losses instantly. Especially in the context of the continuous improvement of industrial automation, high-speed rotating intelligent equipment has increased the potential risk level of dangerous areas, making the traditional safety management methods face severe challenges.

[0003] Traditional personnel entry detection methods mainly rely on manual monitoring or simple sensor detection, and have the following problems: Manual monitoring has low efficiency and is easily affected by factors such as fatigue and inattention, and cannot achieve 24-hour uninterrupted monitoring; Sensor detection has low accuracy, and traditional devices such as infrared sensors and laser sensors are easily interfered by the environment, with high false alarm rates and missed detection rates. In recent years, although detection technologies based on computer vision have been applied to some extent, single-camera solutions still have limitations in accuracy in complex scenarios such as low illuminance and occlusion, and are prone to false alarms and missed detections. First of all, single-camera solutions are greatly affected by environmental light and are prone to false alarms or missed detections in low-illuminance or strong-light environments; Secondly, traditional infrared sensors cannot provide accurate personnel position information and are easily interfered by high-temperature equipment; These limitations make it difficult for existing systems to achieve stable and accurate personnel detection in complex industrial scenarios, so a security solution is needed. Summary of the Invention

[0004] The purpose of the present invention is to provide a dual-camera personnel security detection method to solve the problems raised in the above background technique.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A dual-camera personnel security detection method, the specific steps are as follows:

[0006] Step 1: Dual-camera deployment

[0007] Install an RGB camera and a thermal infrared camera to ensure that the fields of view of both cover the same dangerous area, and adjust the focal length, angle and position of the dual-camera to ensure the synchronization and spatial consistency of image acquisition;

[0008] Dual-camera calibration: Through the checkerboard calibration method of the calibration tool, the spatial coordinate alignment of the dual cameras is completed, the mapping relationship between the RGB image and the thermal infrared image is established, and at the same time, a dynamic recalibration mechanism is adopted to automatically detect the camera offset every three months to ensure long-term stability;

[0009] Step 2: Data collection and preprocessing

[0010] Synchronized image acquisition: Through the hardware trigger or software synchronization mechanism, ensure that the RGB camera and the thermal infrared camera collect images simultaneously, and set the acquisition frequency to 10 frames / s to meet the real-time requirements;

[0011] RGB image preprocessing: Normalize the RGB image, adjust the brightness and contrast, reduce the influence of light changes, and dynamically select the filtering algorithm according to the light intensity to eliminate noise;

[0012] Thermal infrared image preprocessing: Automatically adjust the threshold range according to the ambient temperature, perform temperature threshold segmentation on the thermal infrared image, lock the high-temperature area of the human body, and apply histogram equalization to enhance the local contrast and highlight the human body contour;

[0013] Perform morphological processing: First, use the GrabCut algorithm to pre-segment the image to reduce noise interference, then perform opening operation, erode first and then dilate to eliminate noise points, then perform closing operation, dilate first and then erode to fill the holes inside the target, and finally adopt contour refinement processing to improve the integrity of the target;

[0014] Step 3: Object detection and feature extraction

[0015] RGB image object detection: Use ResNet as the backbone network to extract features, obtain the feature maps of three layers c3, c4, and c5, with sizes of 1 / 8, 1 / 16, and 1 / 32 of the original image respectively. The c3, c4, and c5 are fused with multi-scale features through the FPN feature pyramid network structure to obtain the feature maps p3, p4, p5, p6, and p7. Preset multiple anchor boxes on p3 to p7, predict the target category through the classification sub-network, predict the position offset of the detection box through the regression sub-network, and then combine the preset anchor boxes to obtain the coordinate prediction on the multi-scale feature layers;

[0016] Post-processing: Use non-maximum suppression (NMS) to filter redundant detection boxes to obtain the category and coordinates of the final personnel detection results;

[0017] Personnel detection in thermal infrared images: Process the image through the Gaussian mixture model, use the processed thermal infrared image, extract the contour of the high-temperature area, generate a personnel detection rectangle box, and verify the consistency of consecutive frames through optical flow tracking after the detection box is output to filter out instantaneous false detections and output coordinate information;

[0018] Step 4: Bimodal Data Fusion and Logical Judgment

[0019] Coordinate Alignment: Based on the spatial coordinate mapping model of dual cameras, align the thermal infrared detection frame with the RGB detection frame;

[0020] Intersection over Union (IoU) Matching: Calculate the overlap rate Intersection over Union (IoU) between the thermal infrared detection frame and the RGB detection frame, and filter out the detection frames with an overlap rate exceeding the preset threshold, which are confirmed as valid personnel targets;

[0021] Adaptive Strategy Selection: In well-lit scenarios, adopt the IoU intersection strategy, requiring a high overlap between the bimodal detection results; in low-light environments, give priority to using the thermal infrared detection results; in the initial stage of model deployment or when there is little heat source interference, adopt the union strategy or a single thermal infrared detection result;

[0022] Step 5: Alarm and Shutdown Control

[0023] Person Entering Judgment: When the logical judgment module confirms that a person enters the dangerous area, trigger an alarm signal, activate the audible and visual alarm device, and send an alarm message to the monitoring center or the mobile terminal of the management personnel;

[0024] Equipment Shutdown Control: Send a shutdown instruction through the industrial control interface PLC to stop the operation of the equipment in the dangerous area.

[0025] Preferably, the RGB camera described in Step 1 enables the HDR mode in strong light or backlight scenarios to avoid overexposure or underexposure.

[0026] Preferably, the dynamic selection filtering algorithm described in Step 2 is used to eliminate noise. Gaussian filtering is adopted in strong light conditions, and non-local means denoising is used in low light.

[0027] Preferably, for the pre-segmentation of the GrabCut algorithm described in Step 2, an automatic initialization segmentation method based on graph cuts is adopted before the opening operation. First, an initial foreground / background mask is established through the temperature threshold result. The foreground is the connected region with a temperature range of 30 - 37°C, specifically subject to the preprocessing threshold range of the thermal infrared image. The background includes the buffer zone. Subsequently, a 5-component GMM model is constructed for iterative optimization. Each iteration includes three steps: GMM parameter re-estimation, graph structure construction, and minimum cut solution. The typical configuration is 3 to 5 iterations and a spatial weight coefficient of 0.5. Finally, a preprocessing result with accurate edges and less noise is output, which significantly improves the edge accuracy.

[0028] Preferably, in step two, the contour refinement process improves the target integrity through multi-stage optimization after closing operation. First, the Suzuki85 algorithm is used to extract the closed contour and filter out small regions with an area < 100 pixels. Then, the Douglas-Peucker algorithm is used for polygon approximation with ε = 0.01 × perimeter. Next, the abrupt change points are smoothed based on curvature calculation. Finally, through sub-pixel corner optimization, the edge positioning accuracy reaches 0.1 pixel level.

[0029] Preferably, in step three, the Gaussian mixture model heat source classification is achieved by establishing a three-dimensional feature vector based on temperature value, local temperature variance, and neighborhood gradient intensity, and using the EM algorithm to fit a GMM model containing two components of human body and interference source. The temperature range of the human body component is set to 32 - 37 °C, specifically subject to the preprocessing threshold range of the thermal infrared image. By calculating the posterior probability of the candidate heat source region, when P(human|x) > 0.7, it is determined as a human body, realizing accurate heat source classification. The feature dimension of this method uses 3 dimensions, temperature + texture + gradient, the number of GMM components is 2, binary classification of human body / non-human body, the temperature threshold is controlled within the range of 32 - 37 °C, specifically subject to the preprocessing threshold range of the thermal infrared image, and the probability threshold is 0.7. Compared with the traditional threshold method, it can reduce the false alarm rate through texture analysis, while maintaining the processing speed and significantly improving the detection rate of small targets.

[0030] Preferably, in step three, for the optical flow tracking verification, the GoodFeaturesToTrack algorithm is used to select up to 20 corner features with a quality level of 0.01 and a minimum distance of 10 pixels within the detection frame. Through the Pyramid Lucas-Kanade algorithm, with a window size of 15×15 and 3 pyramid levels, the motion vectors of consecutive frames are calculated, the displacement amplitude is analyzed, the effective range is 0.2 - 5 pixels / frame corresponding to the human walking speed, and the variance of the motion direction, with a threshold of 30°. At the same time, the change rate of IoU between adjacent frames is calculated, and ΔIoU > 0.3 is determined as an abnormal mutation. For the targets that appear instantaneously, three types of abnormal situations, namely, the trajectory length < 3 frames, non-physiological motion speed > 5 pixels / frame or direction mutation > 90°, and heat source diffusion ΔIoU > 0.3 for two consecutive frames, are filtered. Finally, the detection results verified by kinematics are output.

[0031] Preferably, for the automatic adjustment of the threshold range in step two, it is 30 - 40 °C in summer and 25 - 37 °C in winter.

[0032] Preferably, the acoustic and optical warning device in step five includes a buzzer, a voice alarm, and a warning light.

[0033] The beneficial effects of the present invention are as follows:

[0034] Through the collaborative deployment and precise calibration of RGB cameras and thermal infrared cameras, all-weather and full-scene reliable monitoring is achieved, effectively overcoming the detection blind spots of single sensors in complex environments; the Retinanet object detection algorithm of deep learning can use a low threshold to greatly improve the detection rate of human targets in images, and combined with thermal infrared imaging, it further improves the accuracy and robustness of detection; and by processing images with the Gaussian mixture model, whether it is light changes, temperature fluctuations or equipment interference, stable performance can be maintained, demonstrating excellent versatility and scalability; through alarm and shutdown control, personnel safety protection is achieved, overall improving the accuracy, real-time performance and adaptability of personnel detection in dangerous areas, while reducing the false alarm rate and missed alarm rate, and ensuring safe operation in industrial environments. Brief Description of the Drawings

[0035] Figure 1 It is the network structure diagram of the Retinanet of the present invention. Detailed Embodiment

[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0037] As Figure 1 shown, the embodiment of the present invention provides a dual-camera personnel security detection method, and the specific steps are as follows:

[0038] Step 1: Dual-camera deployment

[0039] Install an RGB camera and a thermal infrared camera to ensure that the fields of view of both cover the same dangerous area. Adjust the focal length, angle and position of the dual cameras to ensure the synchronization and spatial consistency of image acquisition;

[0040] Dual-camera calibration: Through the calibration tool checkerboard calibration method, complete the spatial coordinate alignment of the dual cameras, establish the mapping relationship between the RGB image and the thermal infrared image, and at the same time adopt a dynamic re-calibration mechanism to automatically detect camera offsets every 3 months to ensure long-term stability;

[0041] Step 2: Data acquisition and preprocessing

[0042] Image synchronous acquisition: Through the hardware trigger or software synchronization mechanism, ensure that the RGB camera and the thermal infrared camera acquire images simultaneously, and set the acquisition frequency to 10 frames / s to meet the real-time requirement;

[0043] RGB Image Preprocessing: Normalize the RGB image, adjust the brightness and contrast, reduce the impact of illumination changes, and dynamically select a filtering algorithm based on the illumination intensity to eliminate noise;

[0044] Thermal Infrared Image Preprocessing: Automatically adjust the threshold range according to the environmental temperature, perform temperature threshold segmentation on the thermal infrared image, lock the high-temperature area of the human body, and apply histogram equalization to enhance the local contrast and highlight the human body contour;

[0045] Perform Morphological Processing: First, use the GrabCut algorithm to pre-segment the image to reduce noise interference, then perform an opening operation, erode first and then dilate to eliminate noise points, then perform a closing operation, dilate first and then erode to fill the internal holes of the target, and finally adopt contour refinement processing to improve the integrity of the target;

[0046] Step Three: Object Detection and Feature Extraction

[0047] RGB Image Object Detection: Use ResNet as the backbone network to extract features, obtain three feature maps of c3, c4, and c5, with sizes of 1 / 8, 1 / 16, and 1 / 32 of the original image respectively. c3, c4, and c5 are fused with multi-scale features through the FPN feature pyramid network structure to obtain p3, p4, p5, p6, and p7 feature maps. Preset multiple anchor boxes on p3 to p7, predict the target category through the classification sub-network, predict the position offset of the detection box through the regression sub-network, and then combine the preset anchor boxes to obtain the coordinate prediction on the multi-scale feature layer;

[0048] Post-processing: Use non-maximum suppression (NMS) to filter redundant detection boxes to obtain the category and coordinates of the final personnel detection result;

[0049] Personnel Detection in Thermal Infrared Images: Process the image through the Gaussian mixture model, use the processed thermal infrared image, extract the contour of the high-temperature area, and generate a personnel detection rectangle box. After the detection box is output, verify the consistency of consecutive frames through optical flow tracking, filter out instantaneous false detections, and output coordinate information;

[0050] Step Four: Dual-Modal Data Fusion and Logical Judgment

[0051] Coordinate Alignment: Based on the spatial coordinate mapping model of the dual cameras, align the thermal infrared detection box with the RGB detection box;

[0052] Intersection over Union (IoU) Matching: Calculate the overlap rate (Intersection over Union, IoU) of the thermal infrared detection box and the RGB detection box, screen out the detection boxes with an overlap rate exceeding the preset threshold, and confirm them as valid personnel targets;

[0053] Adaptive strategy selection: In well-lit scenarios, the Intersection over Union (IoU) intersection strategy is adopted, requiring a high overlap between the bimodal detection results; in low-light environments, the thermal infrared detection results are preferentially used; in the initial stage of model deployment or when there is little heat source interference, the union strategy or single thermal infrared detection results are adopted.

[0054] Step Five: Alarm and shutdown control

[0055] Person entry determination: When the logic judgment module confirms that a person enters the dangerous area, an alarm signal is triggered, the audible and visual alarm device is activated, and an alarm message is sent to the monitoring center or the mobile terminal of the management personnel.

[0056] Equipment shutdown control: A shutdown instruction is sent through the industrial control interface PLC to stop the operation of the equipment in the dangerous area.

[0057] A spatial mapping relationship is established through the precise deployment and calibration of the dual cameras to achieve synchronous image acquisition and adaptive preprocessing; the RetinaNet algorithm is used for multi-scale RGB target detection, combined with the intelligent temperature segmentation and morphological optimization of the thermal infrared image to extract the target features of personnel respectively; a bimodal data fusion strategy is adopted, and reliable personnel determination is achieved through coordinate alignment and dynamic IoU matching; finally, the detection mode is adaptively selected according to the environmental light conditions, and alarm and equipment shutdown control are triggered, forming an all-weather security solution that takes into account both accuracy and robustness.

[0058] Among them, in step one, the RGB camera enables the HDR mode in strong light or backlight scenarios to avoid overexposure or underexposure.

[0059] The RGB camera enabled with the HDR mode can intelligently balance the bright and dark areas in the picture, automatically expand the dynamic range in strong light or backlight scenarios, ensure the details of both the highlight area and the shadow area are retained, and avoid the loss of target information caused by exposure imbalance.

[0060] Among them, in step two, the filtering algorithm is dynamically selected to eliminate noise. Gaussian filtering is used in strong light conditions, and non-local means denoising is used in low light.

[0061] By dynamically selecting the filtering algorithm to intelligently adapt to the noise reduction requirements under different lighting conditions, the noise elimination effect is improved, and at the same time, the problem of feature loss caused by over-smoothing is avoided, ensuring the accuracy of subsequent RGB image target detection in various lighting environments, and significantly enhancing the overall robustness of the system.

[0062] Among them, in step two, the GrabCut algorithm pre-segmentation uses an automatic initialization segmentation method based on graph cut before opening operation. First, an initial foreground / background mask is established through the temperature threshold result. The foreground is the connected region of 30 - 37°C, specifically subject to the preprocessing threshold range of the thermal infrared image. The background includes the buffer zone. Subsequently, a 5-component GMM model is constructed for iterative optimization. Each iteration includes three steps: GMM parameter re-estimation, graph structure construction, and minimum cut solution. The typical configuration is 3 to 5 iterations and a spatial weight coefficient of 0.5. Finally, a preprocessing result with accurate edges and less noise is output, with a significant improvement in edge accuracy.

[0063] By using the GrabCut algorithm pre-segmentation for processing before opening operation, the edge accuracy and noise suppression effect of the human target in the thermal infrared image can be significantly improved. Through the bimodal optimization combining the energy minimization framework based on graph cut and temperature threshold initialization, the adhered heat sources that are difficult to eliminate by traditional morphological processing can be realized with a high accurate segmentation rate.

[0064] Among them, in step two, the contour refinement processing improves the target integrity through multi-stage optimization after closing operation. First, the Suzuki85 algorithm is used to extract the closed contour and filter out small regions with an area < 100 pixels. Then, the Douglas-Peucker algorithm is used for polygon approximation with ε = 0.01 × perimeter. Then, the smooth mutation points are calculated based on curvature. Finally, through sub-pixel corner optimization, the edge positioning accuracy reaches 0.1 pixel level.

[0065] By adopting the contour refinement processing, the missed detection rate can be reduced and the occlusion robustness can be improved. Overall, a complete optimization chain from pixel level to sub-pixel level is formed, and the intersection over union (IoU) index of the detection box is improved under the same computing resources, significantly improving the matching accuracy of subsequent bimodal fusion.

[0066] Among them, in step three, the Gaussian mixture model heat source classification establishes a three-dimensional feature vector based on temperature value, local temperature variance, and neighborhood gradient intensity, and uses the EM algorithm to fit a GMM model containing two components: human body and interference source. The temperature range of the human body component is set to 32 - 37°C, specifically subject to the preprocessing threshold range of the thermal infrared image. By calculating the posterior probability of the candidate heat source region, when P(human|x) > 0.7, it is determined as a human body, realizing accurate heat source classification. The feature dimension of this method uses 3 dimensions: temperature + texture + gradient, the number of GMM components is 2, binary classification of human body / non-human body, the temperature threshold is controlled within the range of 32 - 37°C, specifically subject to the preprocessing threshold range of the thermal infrared image, and the probability threshold is 0.7. Compared with the traditional threshold method, it can reduce the false alarm rate through texture analysis while maintaining the processing speed, significantly improving the detection rate of small targets.

[0067] Using the Gaussian Mixture Model (GMM) for heat source classification can effectively model complex heat distribution patterns. By probabilistically mixing multiple Gaussian components, it can accurately describe the essential differences between human heat sources and industrial interference sources, enhancing the ability to distinguish interference sources.

[0068] Among them, in step three, for optical flow tracking verification, the GoodFeaturesToTrack algorithm is used to select up to 20 corner features with a quality level of 0.01 and a minimum distance of 10 pixels within the detection box. Through the Pyramid Lucas-Kanade algorithm, with a window size of 15×15 and 3 pyramid levels, the motion vectors of consecutive frames are calculated, and the displacement amplitude is analyzed. The effective range of 0.2 - 5 pixels / frame corresponds to the walking speed of a human body, and the variance of the motion direction, with a threshold of 30°. At the same time, the change rate of IoU between adjacent frames is calculated. When ΔIoU > 0.3, it is determined as an abnormal mutation. For transiently appearing targets, three types of abnormal situations, namely, a trajectory length < 3 frames, a non-physiological motion speed > 5 pixels / frame or a direction mutation > 90°, and a heat source diffusion with ΔIoU > 0.3 for two consecutive frames, are filtered. Finally, the detection results verified by kinematics are output.

[0069] By using the GoodFeaturesToTrack algorithm, the most discriminative corner features within the detection box can be efficiently extracted, ensuring that the selected feature points have stable tracking performance and noise resistance, effectively eliminating false detection problems caused by random jitter and transient interference.

[0070] Among them, in step two, the threshold range is automatically adjusted to 30 - 40°C in summer and 25 - 37°C in winter.

[0071] Automatically adjusting the temperature threshold range can significantly improve the detection reliability of the system under seasonal temperature differences. By dynamically adapting to environmental temperature changes, it can avoid missed detections in summer due to high temperatures or false alarms in winter due to low temperatures with fixed thresholds, ensuring the accuracy of human heat source detection.

[0072] Among them, in step five, the audible and visual alarm device includes a buzzer, a voice alarm, and a warning light.

[0073] By setting the buzzer and the warning light, a warning system in terms of vision and hearing is formed, and security personnel can communicate with personnel in the dangerous area through the voice alarm.

[0074] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprising", "including" or any other variant thereof are intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements not only includes those elements but also includes other elements not expressly listed, or elements inherent to such process, method, article or device.

[0075] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A dual-camera personnel security detection method, characterized in that The specific steps are as follows: Step 1: Dual-camera deployment Install an RGB camera and a thermal infrared camera, ensure that the fields of view of both cover the same dangerous area, and adjust the focal length, angle, and position of the dual cameras to ensure the synchronization and spatial consistency of image acquisition; Dual-camera calibration: Through the calibration tool checkerboard calibration method, complete the spatial coordinate alignment of the dual cameras, establish the mapping relationship between the RGB image and the thermal infrared image, and at the same time adopt a dynamic recalibration mechanism to automatically detect camera offsets every 3 months to ensure long-term stability; Step 2: Data acquisition and preprocessing Synchronous image acquisition: Through a hardware trigger or software synchronization mechanism, ensure that the RGB camera and the thermal infrared camera acquire images simultaneously, and set the acquisition frequency to 10 frames / s to meet the real-time requirements; RGB image preprocessing: Normalize the RGB image, adjust the brightness and contrast, reduce the influence of light changes, and dynamically select a filtering algorithm according to the light intensity to eliminate noise; Thermal infrared image preprocessing: Automatically adjust the threshold range according to the ambient temperature, perform temperature threshold segmentation on the thermal infrared image, lock the high-temperature area of the human body, and apply histogram equalization to enhance the local contrast and highlight the human body contour; Perform morphological processing: First, use the GrabCut algorithm to pre-segment the image to reduce noise interference, then perform opening operation, erode first and then dilate to eliminate noise points, then perform closing operation, dilate first and then erode to fill the internal holes of the target, and finally adopt contour refinement processing to improve the integrity of the target; Step 3: Object detection and feature extraction RGB image object detection: Use ResNet as the backbone network to extract features, obtain three feature maps of c3, c4, and c5, with sizes of 1 / 8, 1 / 16, and 1 / 32 of the original image respectively. c3, c4, and c5 are fused with multi-scale features through the FPN feature pyramid network structure to obtain p3, p4, p5, p6, and p7 feature maps. Multiple anchor boxes are preset on p3 to p7. The classification sub-network predicts the target category, and the regression sub-network predicts the position offset of the detection box. Combining with the preset anchor boxes, the coordinate prediction on the multi-scale feature layer is obtained; Post-processing: Use non-maximum suppression (NMS) to filter redundant detection boxes to obtain the category and coordinates of the final personnel detection results; Personnel detection in thermal infrared images: Process the image through a Gaussian mixture model, use the processed thermal infrared image, extract the contour of the high-temperature area, and generate a personnel detection rectangle box. After the detection box is output, the consistency of consecutive frames is verified through optical flow tracking to filter out instantaneous false detections and output coordinate information; Step 4: Dual-modal data fusion and logical judgment Coordinate alignment: Based on the spatial coordinate mapping model of the dual cameras, align the thermal infrared detection box with the RGB detection box; Intersection over Union (IoU) matching: Calculate the overlap rate intersection over union (IoU) between the thermal infrared detection box and the RGB detection box, and screen out the detection boxes with an overlap rate exceeding the preset threshold to confirm valid personnel targets; Adaptive strategy selection: In well-lit scenarios, the Intersection over Union (IoU) intersection strategy is adopted, requiring a high overlap between the bimodal detection results; in low-illumination environments, the thermal infrared detection results are preferentially used; in the initial stage of model deployment or when there is little heat source interference, the union strategy or single thermal infrared detection results are adopted. Step Five: Alarm and Shutdown Control Personnel entry determination: When the logic judgment module confirms that personnel enter the dangerous area, an alarm signal is triggered, the audible and visual alarm device is activated, and an alarm message is sent to the monitoring center or the mobile terminal of the management personnel. Equipment shutdown control: A shutdown instruction is sent through the industrial control interface PLC to stop the operation of the equipment in the dangerous area.

2. The dual-camera personnel security detection method according to claim 1, wherein: The RGB camera described in Step One enables the HDR mode in strong light or backlight scenarios to avoid overexposure or underexposure.

3. A dual-camera personnel security detection method according to claim 1, characterized in that: The dynamic selection filtering algorithm described in Step Two is used to eliminate noise. Gaussian filtering is adopted in strong light conditions, and non-local means denoising is used in low light.

4. A dual-camera personnel security detection method according to claim 1, wherein: The GrabCut algorithm pre-segmentation described in Step Two adopts an automatic initialization segmentation method based on graph cuts before opening operation. First, an initial foreground / background mask is established through the temperature threshold result. The foreground is the connected area with a temperature range of 30 - 37 °C, specifically subject to the thermal infrared image preprocessing threshold range. The background includes the buffer zone. Subsequently, a 5-component GMM model is constructed for iterative optimization. Each iteration includes three steps: GMM parameter re-estimation, graph structure construction, and minimum cut solution. The typical configuration is 3 to 5 iterations and a spatial weight coefficient of 0.

5. Finally, a preprocessing result with precise edges and less noise is output, with a significant improvement in edge accuracy.

5. A dual-camera personnel security detection method according to claim 1, characterized in that: The contour refinement processing described in Step Two improves the target integrity through multi-stage optimization after closing operation. First, the Suzuki85 algorithm is used to extract the closed contour and filter out small regions with an area < 100 pixels. Then, the Douglas-Peucker algorithm is used for polygon approximation with ε = 0.01 × perimeter. Then, the smooth mutation points are calculated based on curvature. Finally, sub-pixel corner optimization is performed to make the edge positioning accuracy reach the 0.1 pixel level.

6. A dual-camera personnel security detection method according to claim 1, characterized in that: The Gaussian mixture model heat source classification described in Step Three establishes a three-dimensional feature vector based on temperature value, local temperature variance, and neighborhood gradient intensity, and uses the EM algorithm to fit a GMM model containing two components: human body and interference source. The temperature range of the human body component is set to 32 - 37 °C, specifically subject to the thermal infrared image preprocessing threshold range. By calculating the posterior probability of the candidate heat source region, when P(human|x) > 0.7, it is determined as a human body, realizing precise heat source classification. The feature dimension of this method uses 3 dimensions: temperature + texture + gradient, the number of GMM components is 2, binary classification of human body / non-human body, the temperature threshold is controlled within the range of 32 - 37 °C, specifically subject to the thermal infrared image preprocessing threshold range, and the probability threshold is 0.

7. Compared with the traditional threshold method, it can reduce the false alarm rate through texture analysis while maintaining the processing speed and significantly improving the detection rate of small targets.

7. A dual-camera personnel security detection method according to claim 1, characterized in that: In step 3, for the optical flow tracking verification, a maximum of 20 corner features with a quality level of 0.01 and a minimum distance of 10 pixels are selected within the detection box using the GoodFeaturesToTrack algorithm. Through the Pyramid Lucas-Kanade algorithm, with a window size of 15×15 and 3 pyramid levels, the motion vectors of consecutive frames are calculated, the displacement amplitude is analyzed, the effective range of 0.2 - 5 pixels / frame corresponds to the human walking speed, and the variance of the motion direction is calculated, with a threshold of 30°. At the same time, the change rate of IoU between adjacent frames is calculated, and ΔIoU > 0.3 is determined as an abnormal mutation. For transiently appearing targets, three types of abnormal situations, namely, a trajectory length < 3 frames, a non-physiological motion speed > 5 pixels / frame or a direction mutation > 90°, and a heat source diffusion with ΔIoU > 0.3 for two consecutive frames, are filtered. Finally, the detection results verified by kinematics are output.

8. A dual-camera personnel security detection method according to claim 1, characterized in that: For the automatic adjustment of the threshold range described in step 2, it is 30 - 40°C in summer and 25 - 37°C in winter.

9. A dual-camera personnel security detection method according to claim 1, characterized in that: The acoustic and optical warning device described in step 5 includes a buzzer, a voice alarm, and a warning light.

Images