The Study On End-to-End Downhole Track Detection Based On Deep Learning

Mineral resources are an indispensable material basis for human development and social progress.With the development of science and technology and social economy,the demand and exploitation of mineral resources are also increasing year by year.At the same time,it also brings a series of mine locomotive transportation accidents.The realization of the automatic driving system can effectively reduce the occurrence of transportation accidents,and the track detection is an important part of realizing the mine automatic driving system.The traditional track detection algorithm is limited by the low-level features of manual design,and can not reach a satisfactory level in the accuracy and speed of detection,Moreover,it is very vulnerable to the influence of environmental factors such as light,seeper and cables,resulting in low detection rate and accuracy.In recent years,with the rapid development of artificial intelligence and deep learning technology,relevant theoretical research has been widely used in various fields,such as face recognition,pedestrian detection,etc.,but it has not been involved in the field of downhole track detection.Aiming at the above problems,this paper based on the theory of deep learning and carried out the research of track detection based on deep learning.On the basis of the existing convolutional neural network related algorithm model,an end-to-end downhole track detection algorithm based on deep learning is proposed.It does not need manual design features,and the output data can be directly obtained from the input data.The detection accuracy and efficiency are far superior to the traditional track detection algorithms.The following is the main research content of this article:(1)Research on related theories of deep learning,convolutional neural network and semantic segmentation has been carried out,which provides strong support for the improvement of network model and parameter optimization in chapter 3 and chapter 4.The downhole track video was collected and labeled,and a downhole track training data set was constructed.(2)In view of the special underground environment and the characteristics of track shape,this paper constructs a spatial convolutional neural network suitable for downhole track detection,and realizes the segmentation of the track pixel level.By testing the images in different track scenes and comparing experiments with related algorithms,the adaptability of the model in various environments is tested.It is provedthat the effectiveness of the deep learning detection model compared with the traditional manual design features is verified.The robustness of the model in complex environments such as weak illumination and severe occlusion is verified.The track detection model in this chapter can effectively resist the interference of various complex external environments and has good adaptability in various track shape detection.The average track detection accuracy can reach about 97.29%,with high detection accuracy and good robustness.(3)In order to further meet the real-time application of track detection at high resolution,based on the bilateral segmentation network,we propose to obtain the context path module,the boundary refinement module and the feature fusion module in the original network.Larger receptive pyramid attention module,and channel attention fusion module for integrated spatial path module and pyramid attention module features.The effectiveness of the proposed module is verified by relevant experiments.The network model can extract the track area efficiently and with high quality,and realize the track detection scheme with real-time and accuracy.The detection rate can reach50 FPS,and the detection accuracy can reach 86.79%.