An Improved KCF For Pedestrian Tracking Based On Pedestrian Detection And Deep Feature

Aiming at the fact that the existing pedestrian tracking algorithm has no scale adaptive capability and cannot effectively solve the problem of occlusion,an improved KCF model,IKCFBPDDF(Improved KCF Based on Pedestrian Detection and Deep Feature)is designed.By introducing a neural network that can extract deep features and an improved YOLOv3 target detection network with higher accuracy,this model realizes the scale adaptation for KCF and effectively solves the occlusion problem of target tracking,largely improving the tracking success rate.The main work of the thesis is as follows:(1)An improved YOLOv3 pedestrian detection algorithm is designed.First,we replace the NMS(non-maximum suppression)used by YOLOv3 with Soft-NMS to improve the accuracy of pedestrian detection.Then,by adding a retrieval algorithm,the detection box that was removed by mistake is retrieved to restore the lost target box,thereby further improving the accuracy.Experiments on the PASCAL-VOC data set show that YOLOv3,which is improved by the Soft-NMS and retrieval algorithm,improves the accuracy by 3.1% compared to the original algorithm,while the running speed is not significantly reduced.(2)A tracking method based on the fusion measurement method of deep feature and scale adaptive KCF is designed.The improved YOLOv3 is used for pedestrian detection,and the image information of the pedestrian target is used as a new template of KCF to solve the problem of scale changes.When the HOG feature is not enough to distinguish pedestrians,the accurate location of the pedestrian target is determined by fusing the deep features.When pedestrians are blocked and the KCF target is lost,the last deep feature before the occlusion is compared with the deep features of all pedestrians recognized by YOLOv3 after the occlusion disappears,and the location of the pedestrian target is re-determined based on the similarity.Compared with the original KCF,the improved algorithm improves the pedestrian tracking success rate in complex situations by 36%.Compared with the current mainstream Siam RPN and Dasiam RPN pedestrian tracking models,the improved algorithm also has a small increase in the tracking success rate.