Mine Pedestrian Detection with Deep Learning Network of Parallel Feature Transfer

Mine pedestrian detection is one of the key technologies for realizing un-manned locomotives.The traditional visual feature extraction algorithm cannot effectively address the complex mine roadway environment and poor lighting.This paper proposes a deep learning network for mine pedestrian detection based on parallel feature transfer to ensure high accuracy and strong real-time performance.First,we describe the structure of the network,including the parallel pedestrian identification module,pedestrian location module and the feature transfer block between them.The pedestrian identification module roughly adjusts the position and size of the anchor boxes and filters the negative anchor boxes.The pedestrian location module further improves the regression accuracy and gives the prediction results.Feature transfer block transmits the features to better connect the two modules.Second,we adopt data set amplification,data enhancement and hard negative mining to optimize the training process.Finally,we get the experimental data from the Taoyuan and Xinji mines in Anhui province.The experimental data show that the proposed algorithm can still maintain 63.4%AP at a rereal-time processing rate of 37 FPS.Compared with the YOLOv1,the AP increases by 9.2%.Compared with the M2Det,the calculation efficiency increases by 22 FPS.