Research And Design On Pedestrian Detection System Under The Mine Based On DCNN

The coal industry occupies a pivotal position of energy in our country,coal mine safety,especially under mine safety is the most important in it.Currently,for detections of under mine workers,most coal enterprises use equipped personal positioning systems,etc.These technologies can locate and recognize effectively,however,they cannot evade the situation of replacing card or illegal substitution,their intelligent level and precision are still low.Particularly when monitoring people are neglecting,there will be a huge potential safety hazard.On account of this,the paper combines the applicaions in video image recgnition of DCNN and industrial video surveillance system that have equipped in most coal enterprises raises a pedstrian detection system under mine based on DCNN.For improving speed,the paper uses YOLO object detection system and ameliorate it direct at the special environment under the mine and finally use the Java Web to do a simple achievement of pedestrian detection system under the mine based on revised YOLO.The paper introduces theory of convolutional neural networks upon neural networks,and then introduce the deep learning networks.Detailedly explains the YOLO object detection system's network structure and proceesion of detection,and makes a deep analysis of the shortage of the YOLO that the precision of it is not good.Second the paper improves original YOLO system's training set and network structure aimed at environment of poor quality of video,dull background,single-target detection,and so on under the mine.The paper has re-make training set by surveillance video and modify the network structure by thought of combining representation information of superfical layer and semantic information of deep layer,put the additon of the eighth layer and final layer as the final output of the whole network,and make 3 programmes based on that.First is convolution first and pooling,second is pooling first and convolution,third is make the output of final layer to deconvolute first and adding the eighth layer's output as the final output.The paper uses Caffe framework to experiment and analyze them and chose second scheme as the final one.The essay proves that the performance of modified YOLO is better than original one.Finally the techniques of JavaEE are used to build the pedstrian detection system,the system includes system management,privilege management,detection management,attentance mangement and device management,and use the lab video to do some simulation experiment to test analysis and functional vertificaion.illstrating feasibility of the system.Through the experiment in this paper,it can be seen that the improved YOLO algorithm has better detection effect on the detection of special environment under mine.