Crowded Pedestrian Detection For High-speed Railway Platform Based On Deep Convolutional Networks

As a key technology in high-speed railway operation and maintenance,pedestrian detection plays a vital role in ensuring the safety of high-speed railway station.However,the high-speed railway platform has a large pedestrian flow,crowd pedestrian distribution,large pedestrian scale changes,and serious mutual occlusion,which leads to the problems of low accuracy and high miss rate of the existing detection methods.For this reason,this thesis puts forward the topic of crowd pedestrian detection in high-speed railway platform based on deep convolution network.From the aspects of constructing high-speed railway platform pedestrian dataset,designing small-scale and occlusion detection algorithms,the research on high-speed railway platform crowd pedestrian detection technology has important and practical value to improve the security and stability of high-speed railway system.The main contributions of this thesis are as follows:(1)A high-speed rail platform dataset,named HUST,is constructed for crowd pedestrian detection.There are 2637 images in the dataset,including 33473 pedestrian targets,with an average of 12.7 pedestrians per image.The dataset is collected from more than ten high-speed rail platforms,and covers crowd pedestrian scenes of various scales and different degrees of occlusion.We found that there are pedestrians not detected on such datasets of crowd scenes when using the classical deep learning detection methods,such as Faster RCNN,SSD and Retina Net.(2)In order to solve the problem of high missing rate of small-scale pedestrian detection in the process of variable scale multi-scale pedestrian detection,we propose a dense connected feature pyramid network,named DCFPN.To fuse the depth semantic information and shallow texture information,the features of different scale are connected with learnable weight.At the same time,based on the relatively stable aspect ratio of pedestrians,K-means clustering algorithm is used to estimate the aspect ratio of pedestrians,which makes the anchor suitable for pedestrian detection.Pedestrian detection experiments are carried out in public dataset and high-speed rail platform dataset respectively.Compared with FPN,the AP value of DCFPN improves 1.1%and 1.9 respectively,and the MR-2 value of heavy reduces 1.2%and 1.4%respectively.(3)In order to solve the problem of low precision and high miss rate in crowd pedestrian detection,we propose a pedestrian detector,named GA RCNN,which based on the guided attention mechanism.Our method includes target frame guided attention module(BGAM),occlusion prediction branch(OPB)and occlusion oriented non-maximum suppression(Occ NMS).The BGAM module uses the target frame to guide the generation of spatial attention feature map to enhance the feature and improve the average precision of pedestrian;the OPB branch is used to predict the occlusion rate of the pedestrian frame,and then Occ NMS dynamically sets different suppression thresholds according to the occlusion rate,so that the bounding box of pedestrian with heavy occlusion can be retained.Experiments are conducted in public dataset and high-speed rail platform dataset respectively.Compared with Rep Loss,MGAN and other pedestrian detection algorithms,the AP value of GA RCNN improves 0.6% and 0.9 respectively,and the MR-2 value of heavy occluded pedestrian reduces 1.1% and 1.2% respectively.In this thesis,pedestrian detection algorithm and experimental research are carried out in the above three aspects.The experimental results show that the proposed method can effectively improve the precision of crowd pedestrian detection and reduces the miss rate of dense pedestrian detection,which can meet the requirements of crowd pedestrian detection application in high-speed rail platform scene,airport,square and other public places.