Research On Pedestrian Detection And Re-identification In Unmanned Driving

With the development of artificial intelligence,unmanned driving,aiming to safe,efficient,comfortable and energy-saving driving,is in a golden times.Safe driving depends not only on the propulsion of related policies,but also on sufficient technical support.Pedestrian detection is very important,and now still focuses on accuracy,real-time performance and continuity during cross-camera detection.Specifically,the main contributions of our work are listed as follows.Firstly,pedestrians are detected in a single camera.Three object detection frameworks based on region proposals,R-CNN,Fast R-CNN and Faster R-CNN,and two frameworks based on global regression,YOLO and YOLOv2,are analyzed.For its high precision and speed,YOLOv2 is selected and improved by using the scales and aspect ratios criteria.Prior knowledge obtained by K-means clustering on training set can help picking better anchor boxes during predicting,and then a map to real scale is carried out to correct the small offsets of output anchor boxes.Furthermore,classification and localization error are calculated by using multi-task loss function for the algorithm judgment.Experiments in KITTI show that the modified YOLOv2 improves the accuracy by 7.1% on average.Secondly,the tracking algorithm is applied to speed up pedestrian detection.A discriminative tracking algorithm,SiamFC,is adopted because of its high speed and accuracy,and the logistic loss function is used in error calculation.Since similarity will be calculated for several times in different locations during tracking,loss function id redefined.In OTB,the proposed algorithm achieves the success rate of 61.8% exceeding the best algorithm by 10%,and the tracking speed of 86 FPS.Finally,pedestrian re-identification is carried out across different cameras.On the study of two re-identification methods based on verification models and identification models,a new algorithm is proposed.ResNet-50 is adopted to extract features,then match them by square layer,and estimate similarity lastly.The network is composed of both verification and identification parts,which is trained jointly to find optimal solution.Experiments in Market1501 show that the mAP of single query and multi query are 62.30% and 72.55%,respectively.This article provides a solution for pedestrian detection and re-identification in unmanned driving.Testing experiments in relevant datasets show that the proposed algorithms achieve high accuracy,speed and continuity in pedestrian detection.To some extent,this work will promote the successful application of pedestrian detection in unmanned driving.