Underground Target Detection Method Based On YOLO

With the rapid development of computer science,computer vision technology is playing an increasingly important role in daily life and production.More and more industries are applying the low-cost technology of computer vision.Among them,the mining industry is paying more and more attention to the development of smart mining concepts with "intelligence and safety" as the core,computer vision technology in the construction of intelligent mines shine.However,the underground environment is special,the hardware equipment is limited,the traditional target detection method has poor real-time performance and low detection accuracy,so that it cannot be applied in the special underground environment.The existing deep learning target detection method has a relatively large model volume.Therefore,the high-precision,small volume and excellent real-time target detection methods are urgently needed in the fields of intelligent mine construction and automatic coal mining.This paper is based on the YOLOv3 target detection method,starts from the difficulty of deploying large and poor real-time models in special environments such as coal mines.This paper proposes two methods: P-YOLO and PG-YOLO,which are combined with sparse pruning and generalized intersection over union.The main work of this paper is as follows:(1)The YOLOv3 target detection method has powerful performance,but its model volume is large,storage and calculation costs are high,and it is difficult to deploy on restricted hardware platforms such as underground mines.This paper proposes the PYOLO method based on the YOLOv3 target detection method combined with channel pruning.This method arranges the convolution channels according to the weights through sparse training,and deletes redundant convolution channels,thereby achieving the purpose of reducing the model volume and simplifying the network parameters.The experimental results show that the P-YOLO method compresses the volume of the model well and greatly reduces the parameters of the network model.In comparison with the R-CNN series,YOLOv1 and YOLOv3 methods,the P-YOLO method has the best model volume and detection speed,and the detection accuracy of the model is also relatively excellent.(2)P-YOLO method uses the intersection over union of the prediction box and the real box to calculate the position loss function,but the intersection over union itself cannot reflect the true overlap of the prediction box and the real box.When the boxes do not intersect,the loss function is difficult to converge.This paper improves the PYOLO method and proposes the PG-YOLO method,which uses the generalized intersection over union loss function instead of the original intersection over union loss function.The experimental results show that the model accuracy of the PG-YOLO method has been improved by 0.8%.In comparison with the R-CNN series and the YOLO method,the PG-YOLO method has the best model accuracy,model volume,and detection speed.(3)Manually mark the establishment of a small coal mine underground data set.Through the training of underground coal mine datasets,the PG-YOLO method improves the detection effect of targets in underground coal mines and effectively reduces the false detections,which lays the foundation for the actual deployment of this method in underground coal mines.There are 35 figures,14 tables and 85 references in this paper.