The Study On Mine Pedestrian Detection Based On Parallel Feature Transfer Of Deep Learning Network

Mineral resources play an extremely important role in the development of national economy and society.Automated production means such as unmanned mining machinery and unmanned harvesters are helpful to reduce the occurrence of mine accidents,while computer vision is an effective technical means to support automated production,in which mine pedestrian detection is the basic task of computer vision.The traditional pedestrian detection algorithm based on feature extraction has many disadvantages in the complex and changeable scene of mine roadway and bad lighting,which cannot guarantee the accuracy of detection results.Fortunately,the development of artificial intelligence brings the dawn of pedestrian detection in the mine.Deep neural network and deep learning algorithm have been applied in various fields in the industrial field,but there is little research in the field of pedestrian detection in the mine.Aiming at the above problems,a mine pedestrian detection algorithm based on parallel feature transfer deep learning network is proposed in this thesis.The main research contents are as follows:(1)A non-pre-trained underground pedestrian detection network automatically generated based on anchor box is proposed to achieve the goal of improving the structural adaptability of the pedestrian detection network,reducing the workload of pre-training and reducing the risk of "negative migration".This network finally named Gas Net.Gas Net is mainly composed of non-pre-trained backbone network and parallel processing detection network.Parallel processing detection network includes anchor box generation module composed of anchor box location prediction and anchor box shape prediction and feature map adaptation module.Through the experiment verification,the network detection accuracy reaches 64.7,the detection performance is good.(2)The mine pedestrian detection network combined with parallel feature transmission was proposed to improve the detection speed while maintaining the accuracy,named as Pft Net.Pft Net is composed of pedestrian identification module,pedestrian location module and feature transfer block.The proposed network model can achieve the real-time detection rate of 37 frames /s and maintain the detection accuracy of 63.4%.Moreover,experiments are carried out on UMP2018 and public pedestrian data set to verify the robustness of the network.Finally,the thesis is summarized and the future research work is prospected.