Research And Achievement Of Efficient People Counting Algorithm Based On Deep Learning

With the development of economy and social security system,all walks of life have more and more needs for people counting and analysis in the region.In terms of the security monitor,it can protect the classified information and property security of country and enterprise.In terms of dispatching,it can optimize the construction location of public facilities to save resources.In terms of personal property security,it can avoid security risks caused by excessive population density.People-counting Methods based on computer vision provide a feasible solution.It mainly extracts the pedestrian features in images and analysis the number of people through the features.However,most people counting methods nowadays could not achieve a trade-off on speed,precision and generalization ability.Normally,with the improvement of precision and generalization,there is a sharp decline on speed performance.Aiming at this problem,we propose a people counting strategy combining object detection and object tracking to improve the accuracy,efficiency and generalization ability of multiple scenes.The main work and contributions of the paper are as follows:(1)Based on the development of related works on people counting and object detection currently,we analysis the main ideas and problems of these works.We adopt the strategy of head-shoulder detection combined with multi-object tracking.The head-shoulder detection is based on SSD algorithm,and the tracker is based on SORT algorithm.After tracking,we obtain the trajectories of all people to determine the number of people.(2)Firstly,we optimize the backbone of the head-shoulder detection.The new backbone reduces the size of feature maps rapidly to improve the efficiency of feature extraction.Then,we propose inception structure to merge features of different layers.By this way,the backbone network could extract more semantic features and remain the detailed information.In addition,we use the feature of dilated convolution that expanding receptive field without increasing computation to expand the receptive field associated with each target.In this case,head-shoulder detection could use the context feature of pedestrians to assist in improving the classification performance of hard examples.(3)Secondly,we optimize the matching strategy between the prior boxes and groundtruth while training.The matching strategy directly affects the quantity and quality of the prior boxes matched with groundtruth.Furthermore,it affects the performance of recall and precision.We improve the matching strategy so that the number of prior boxes matched with small groundtruth increases a lot meanwhile remaining the quality of these prior boxes.By this way,the recall of small targets increases a lot.(4)Thirdly,we optimize the classification and regression loss.The improved repulsion loss suppress the influence caused by NMS algorithm when people occlude with each other.Focal loss enhance the representation power of hard examples which increasing the classification performance.(5)Under the situation that video decoding interfered,SORT algorithm loss the tracking targets.We adopt IoU tracker combined with SORT tracker to improving the generalization performance.We collect a head-shoulder dataset to verify the effectiveness of each module in the algorithm.The comparison experiments verify the effectiveness of each module.Our algorithm proposed in this paper achieves state-of-the-art results on speed,accuracy and generalization ability.