Research On Pedestrian Detection Algorithm Based On Improved End-to-End Deep Learning Framework

With the development of metro in various large and medium-sized cities,it has gradually become the backbone of urban transport capacity,which has put forward higher requirements for urban traffic management,metro dispatching,distribution of transport Resources and public safety.In this context,passenger flow detection and statistics of subway become the key technologies to provide data support,and pedestrian detection is the basic key technology of passenger flow detection.With the development of computing power,the current target detection algorithm based on deep learning has been relatively mature.In order to take into account more kinds of scenarios and targets,the current popular target detection network generally has a deeper network layer.The pedestrian detection in this paper is single target detection,and the feature of head and shoulder is relatively simple.In this case,deep networks are not fully applicable,since there will be large computational redundancy,resulting in a slow detection speed.In addition,single targets are relatively small in the picture in subway pedestrian detection,and there are many pedestrian dense scenes,with pedestrian shielding phenomenon,which will also lead to the decline of expression of common network and the problem of missing detection.Experiments show that the improved end-to-end deep learning framework can effectively improve the detection speed while ensuring the detection accuracy.In this paper,the pedestrian detection algorithm based on end-to-end deep learning framework in the subway environment is mainly aimed at the slow detection speed caused by network redundancy and the missed detection caused by occlusion and the false detection caused by occlusion,and the existing typical end-to-end deep learning framework is simplified and optimized.The specific work is as follows:1.According to the background of this task,the data set of pedestrian images in subway environment was made and divided into training set,verification set and test set.According to the collected data set,K-means++ clustering algorithm is used to cluster the size of anchor frames in the process of network detection,and the number of anchor frames is improved through experiments.The detection accuracy and speed of the improved network model are improved.2.Aiming at the real-time problem in pedestrian detection,the backbone network of YOLOv3,a classic network framework for end-to-end deep learning,was pruned.According to the characteristics of subway pedestrians and the characteristics of single-target detection,the optimal backbone network complexity was found and the target detection network is simplified,so that the detection speed of the model framework can be improved more than three times while ensuring the detection accuracy.3.Aiming at the problem of occlusion in subway pedestrian detection,a YOLOv3 network model framework based on the repulsion loss function is proposed,which is combined with the original loss function to guide the model to converge to a better direction,making the model more suitable for pedestrian detection in dense scenes.After training and testing the network framework on the high-performance computing platform,the model that has been learned can quickly and efficiently detect pedestrians from different scenes and angles in the subway environment.