Research On Pedestrian Detection And Tracking Based On Deep Learning Feature Sharing

Nowadays,with the advent of the artificial intelligence age,there has been a wave of research on unmanned driving systems around the world.The driving safety problem of smart car is the key point in the research.How to ensure the safety of vehicle itself and other participant of traffic environment like pedestrian and other vehicles during the driving of the unmanned vehicle is a serious problem.Therefore,it is necessary for unmanned vehicles to obtain real-time and reliable pedestrian's status information and make pre-judgments in advance.That can avoid threating or harming to the safety of pedestrian.In addition,the performance of tracking depends on feature expression.In response to this problem,this subject studies pedestrian detection,pedestrian tracking and feature expression.The thesis utilizes a detection-based multi-pedestrian tracking framework.In terms of pedestrian detection,the deep convolutional network,Mask RCNN,is modified by MobileNet.On this basis,this thesis replaces the non-maximum suppression method in the original network with the method of soft maximal suppression for further improving the accuracy of detection.The test verification on the MS COCO dataset shows that the modified Mask RCNN model has the same detection accuracy of the original Mask RCNN.And the processing speed of modified Mask RCNN is about 20% higher than the processing speed of Mask RCNN.In order to match the pedestrian targets in the continuous frame pictures,the thesis uses the hungarian algorithm.And the matching cost matrix is constructed by appearance characteristics and motion characteristics of pedestrian.Then the positions of targets in the next frame image are predicted by kalman filter.In aspect of pedestrian characterization in tracking,we consider that deep learning feature has outstanding advantages compared with traditional artificial design features,but the method increases operation time and cost of computing.we directly use 1024-dimensional features of the Mask RCNN as the pedestrian appearance characteristics in the tracking process.The final appearance features are incorporated by color histogram features and 1024-dimensional features.The simulation on the MOT16 dataset shows that the proposed method is feasible.In this paper,the weights of appearance feature and motion feature of pedestrian are adjusted.The results show that the MOTA is 34.3%,and the MOTP is 79.3% when appearance feature weight is set to 0.8.Compared to the reference model,the MOTA of model is increased by 5.3% and the MOTP of model is increased by 0.5%.