Research On Pedestrian Detection Approach Based On Convolutional Neural Network

Pedestrian detection is one of the tasks of computer vision,which is not only the basis for other computer vision tasks such as person search,pedestrian re-identification pedestrian behavior analysis,pedestrian posture estimation,and so on,but also widely used in video surveillance,autonomous driving,intelligent robots,and other fields.There are many challenges in pedestrian detection such as occlusion,variety of posture,illumination,and small pedestrians.Anchor＿free CSP(Center and Scale Prediction)algorithm is proposed,the CSP algorithm has a simple structure.In this paper,it is argued that CSP only predicts the centroid and height of a pedestrian,and that the detector needs to learn additional features to identify pedestrians more accurately.Based on this,firstly,this work proposes multi-task learning to guide pedestrian detection;secondly,this work proposes to let the detector learn prior information to better locate pedestrians.To improve pedestrian detection performance,the main work is as follows:(1)This study proposes an anchor-free pedestrian detection approach(CSPRS),which is based on CSP and combines region Resolution learning and region fine-grained Segmentation learning.The box for the visible part of the pedestrian is smaller than the full body surround box for the pedestrian when the pedestrian is in occlusion,the fullbody bounding box contains part of the pedestrian,other objects,and the background.Based on this,this work proposes two different ways to learn more features of the pedestrian in different occlusion situations.First,this work proposes region resolution learning which learns the pedestrian regions on the input image.Second,this work proposes fine-grained segmentation learning to learn the outline and shape of different parts of pedestrians.The detector learns not only pedestrian location features,but also extra features of the pedestrians.CSPRS perceives the feature of pixels,outlines and shapes of pedestrian within the pedestrian full-body bounding box areas.Region resolution learning and region segmentation learning help the detector to locate pedestrians.Experimental results show that both ways of learning pedestrian features improve performance.This work evaluates our proposed detector CSPRS on the City Persons benchmark,and the experiments show that CSPRS achieves 42.53% on the heavy set on City Persons datasets.(2)This study proposes OCSP which is based on CSP and adopts semantic head center and visible part center to handle occlusion.To deal with the occlusion challenge,OCSP integrates different priors.OCSP predicts semantic head center,visible part center,body center,and scale.There are many occlusion scenes in pedestrian detection,the pedestrian head has identifiable,unique features.In this way,the detector perceives semantic head features to better localize pedestrian centers.Secondly,at different levels of occlusion,the detector senses the centre point of two different areas: the centre point of the visible frame and the centre point of the pedestrian body frame,and the pedestrian visible part centroid assists the detector in locating the pedestrian centroid.Based on this,this work also designs a branch for predicting semantic head center.This work also lets the detector predict visible part boxes.Experiments show that this fusion of prior information improves the performance.On the City Persons dataset,OCSP(w)achieves39.9% on the heavy set,and OCSP* achieves 8.35% on the reasonable set.