Reaearch On High Accuracy Pedestrian Detection

The research of object detecton is currently popular in the field of computer vision,which is aimed to detemine the position and category of some object.As a essential branch of object detection,pedestrian detection has been widely applied to many fields,sunch as security,intelligent driving,unmanned aerial vehicles,and so on.In recent years,deep learning has greatly boosted the development of identification,detection,segmentation,etc.,which result in the solution of problems such as pedestrian posture,complex scene background,the density of pedestrian,pedestrian scale,illumination,etc.Deep learning has significantly improve the accuracy of detection while the miss rate is decreased efficiently,so,based on the background above,pedestrian detection has been deeply researched in this paper,the main works include:Firstly,Faster R-CNN has been one of the most popular algorithm of general object detection,though appling it to pedestrian detection derectly cannot meet the requirement in practice.In order to integrate this algorithm into pedestrian detection,this thesis proposes an algorithm based on deep learning and the gradual strategy of transfer learning:PDA-DLGT.This alogorithm has modified the networks' structure,furthermore,it adopts gradual strategy of transfer learning which uses a hybrid database,INRIA+ETH,as the intermediate dataset.The modified networks in PDA-DLGT result in the better fitting and extracted features of pedestrian data compared to traditional networks.Experiment results show that the gradual strategy of transfer learning can decrease the miss rate by approximately 5%,and PDA-DLGT is with the fairly comparative performance,which is better than JointDeep.Secondly,compared with the general object detection,the specific problems have been analyzed.The crux lie mainly in the sacle of pedestrian and the large complex background.According to these two problems,this thesis proposes the solution,which is an algoritm based oil region proposals network and cascade Boosting forest?PDA-RPNCBF?.This algorithm combines the CNN feature of conv3₃ and conv4₃.In addition,PDA-RPNCBF itroduces the cascade Boosting forests and Bootstrapping to mine the hard-negative example.Specifically,PDA-RPNCBF adopts a complete and progressive method,which will not only ensure the quantity of nagetives,but also the quality of the negative samples,the method can increase the detection accuracy significantly.Experiment results show that the miss rate of PDA-RPNCBF fell to 10.63%.Compared with PDA-DLGT,the miss rate of PDA-RPNCBF has been decreased by nearly 14%,and is superior to the CCF,CompAct-Deep and RPN + BF.Finally,the multi-scale issure of pedestrian detection has been analyzed.Because the scale of the pedestrian varied significantly and small scale is dominated while CNN characteristics of scale invariance is limited,this thesis hypothesizes that the multi-scale problem of deep convolutional neural networks-based pedestrian detection cannot directly adopt the scheme of general object detection.This thesis proposed PDA-MSRPN based on multi-scale region proposal networks and the improvement of PDA-RPNCBF.The algorithm utilizes the cascade Boosting forest of PDA-RPNCBF while the region proposal network is transformed to MSRPN by incorporating the multi-scale factors.MSRPN generates and integrates the region proposals on the multiple network layers by adding micro-network branches.MSRPN can grasp the pedestrian multi-scale information sufficiently due to different branches have different receipt field.Experiment results show that MSRPN can decrease the miss rate by 2.21%compared to RPN.Furthermore,the accuracy of MSRPN has been promoted further by integrating cascade Boosting forests and Bootstrapping.