Abstract:

In order to intelligently monitor the intrusion of personnel entering dangerous areas and the wearing of safety helmets in the fully mechanized mining face of underground coal mines, an improved object detection algorithm based on YOLOv5s was proposed to solve the problem of dust interference and uneven illumination affecting image detection accuracy in monitoring videos. Firstly, in the neck network section, DwConv was used to replace ordinary convolutions, which reduced parameters and computational complexity. Then, an improved PANet was introduced to improve the feature extraction capability, replacing the bounding box Loss function CIOU (Complete-Intersection Over Union) with EIOU to improve the detection accuracy. Finally, the personnel images in the video of the fully mechanized mining face were selected for detection, and monitoring videos of coal mine underground personnel entering and wearing safety helmets were selected as the detection dataset for training and verification. The results show that, compared with the initial YOLOv5s algorithm network model, the number of parameters of the YOLOv5s-DPE algorithm network model is decreased by 14.2%, the number of floating-point arithmetic calculations is decreased by 7.6%, the size of the algorithm network model is decreased by 12.5%, mAP@0.5 is increased to 93.7%, and mAP@0.5:0.95 is increased to 65.8%. The YOLOv5s-DPE model has better detection performance for small targets, with a reduction in false detections and missed detections.

Key words: YOLOv5s, fully mechanized coal face, detection algorithm, depthwise separable convolution (DwConv), efficient intersection over union (EIOU), path aggregation network (PANet)