Research On Pedestrian Detection Algorithm In Dense Scene

At present,pedestrian detection technology has been widely used in various scenes in life,and its purpose is to classify and locate pedestrians in videos or images.However,there are more interferences in application scenarios: mainly environmental factors and human factors.Environmental factors include: cross-situation differences,image acquisition equipment height and angle differences,day and night light problems,weather changes and other factors.Human factors are: the gathering characteristics of pedestrians and the difference in appearance of pedestrians.All of the above factors are the reasons why pedestrian detection algorithms are struggling.The most prominent problem is the clustering characteristics of pedestrians.With the dense scene of pedestrians,the prominent problem is the occlusion,followed by the problem of small object detection.Therefore,the key research direction of this article is to solve the phenomenon of poor performance of pedestrian detection algorithms due to occlusion,and to optimize the feature extraction of pedestrian detection for small targets.This paper divides the occlusion problem in pedestrian detection into inter-class occlusion and intra-class occlusion.The main research content of this paper is composed of the following five parts.The first four parts are targeted optimization for occlusion problems,and the fifth part is optimized design for small target detection.1.Multi-branch model and visable learning strategy to deal with the problem of occlusion between classes.In this paper,a method of combining multi-branch model and explicit learning strategy is proposed for inter-class occlusion.The visable learning strategy can use the visible part of pedestrian annotations to guide the learning process of the model,so that it pays more attention to the visible part of the pedestrian characteristics,and can effectively solve the problem of feature missing between classes.Compared with the single-branch model,the multi-branch model has additional branches.This branch can adopt an explicit learning strategy,which effectively handles the problem of inter-class occlusion while only adding a small computational cost.2.The multi-branch model and multi-instance prediction mechanism deal with the problem of intra-class occlusion.In this paper,a multi-branch model and a multi-instance prediction mechanism are combined for intra-class occlusion.Because the single-branch model has only one prediction branch,it can only achieve single-instance prediction,that is,one candidate frame corresponds to one pedestrian.The multi-branch model has multiple detection heads,and a candidate frame can generate multiple sets of detection results.In the case of intra-class occlusion,multi-instance prediction can effectively improve the missed detection.3.The soft-NMS algorithm handles the problem of missed detection under occlusion.In this paper,aiming at the missed detection problem caused by the non-maximum suppression algorithm of the greedy mechanism,the soft non-maximum suppression algorithm is used to replace it.The recall rate has been improved,and the average accuracy has also been improved.4.Guided data enhancement methods solve the problems of fewer occlusion samples and lack of data enhancement methods for occlusion problems.Aiming at the current situation where the number of occlusion samples in the pedestrian data set is small,this paper designs a well-guided data enhancement method.During training,the pedestrian marking frame is divided into head,left torso,right torso,left leg,and right leg.Five parts,choosing one part for occlusion,effectively solves the problem of fewer occluded samples.5.Use feature fusion and parallel structure backbone network to solve the problem of small target detection.This paper proposes an optimization method for the backbone network of the model,and uses HRNet with a large number of feature fusion to replace the original VGG16 network of Faster-RCNN,which achieves better results in average accuracy and loss rate.Based on the above five improvements,we designed multiple sets of comparative experiments on the public benchmark pedestrian data sets Caltech,City Persons,Crowd Human,and the subway pedestrian data sets collected and labeled at various subway stations in China.In the above four data sets,the method designed in this paper has achieved a significant drop in the loss rate and achieved a good improvement in the average accuracy,which verifies the effectiveness of the design in this paper.