Intelligent recognition method for personnel intrusion hazardous area in fully mechanized mining face

To address the problems of low accuracy of video AI recognition of personnel intrusion hazardous areas in fully mechanized mining face caused by factors such as variable personnel scales, and dynamic changes of hazardous areas, an intelligent recognition method for personnel intrusion hazardous areas of fully mechanized mining face based on RSCA-YOLOv8s and automatic division of hazardous areas is proposed. To address the problem of low accuracy of personnel recognition in fully mechanized mining face, the RFAConv-SE (Squeeze-and-Excitation with Receptive-Field Attention Convolution) and CCNet (Criss-Cross Attention Network) attention modules are introduces on the basis of the YOLOv8s model to improve the capture ability of the model for global and contextual information in complex background images. The multi-scale and small target personnel feature extraction ability of the model is improved by fusing the Res2Net network through the C2f module. The adaptive fusion ability of the model for multi-scale personnel features is enhanced through the improved SPC-ASFF (Adaptive Structure Feature Fusion with Sub-Pixel Convolution layer). To address the problem of dynamic changes of hazardous areas within the field of view caused by the dynamic changes of the camera on fully mechanized mining face following the hydraulic support, an automatic division method of hazardous areas based on the extraction of key feature points of landmark targets such as the guard plate and coal baffle plate is proposed. To address the problems of irregular changes of hazardous areas and the difficulty of parameter setting of the judgment method based on overlap degree, a precise recognition method for personnel intrusion hazardous areas based on the ray method to determine the positional relationship between the pixel coordinates of personnel and hazardous areas is proposed. Through ablation experiments, comparison experiments between RSCA-YOLOv8s and methods such as YOLOv5s and YOLOv8-SPDConv, as well as test experiments on automatic division of seven groups of multi-scene hazardous areas and recognition of five groups of personnel intrusion into hazardous areas in fully mechanized mining faces,the results show that the accuracy of the personnel recognition method of RSCA-YOLOv8s is higher, reaching 97.2%, which is 1.1% higher than the baseline model mAP@0.5 and 2.5% higher than mAP@0.5:0.95. It has more accurate recognition ability and higher recognition accuracy for small target personnel. The average accuracy of the automatic division of hazardous areas by this method is 97.285% and the discrimination accuracy of personnel intrusion hazardous areas is more than 98%.