Research On Video-based Pedestrian Tracking Method

Pedestrian tracking is one of the hot topics in machine vision field.By locating pedestrians and acquiring their tracks,the continuous tracking of pedestrian targets is realized.Due to the interference of fast motion,occlusion and other factors,pedestrian tracking algorithms have problems such as low accuracy,poor robustness and others in practical application.Therefore,the pedestrian tracking method based on video is of much application value.This thesis focuses on the video-based pedestrian tracking method,and researches on pedestrian detection,single-target pedestrian tracking,multi-target pedestrian tracking.The specific work is as follows:In terms of pedestrian detection,this thesis discusses the YOLO series of detection methods.According to the requirements of real-time and accuracy in the actual scene,a pedestrian detection algorithm based on YOLOv3 architecture is implemented.The Caltech dateset is used to test the efficiency and results show that the pedestrian detection method based on YOLOv3 achieves a balance between real-time and accuracy,which lays the foundation for the following research of pedestrian tracking method.In terms of single-target pedestrian tracking,aiming at the problem of low robustness and poor generalization ability of the object tracking method based on Siamese Region Proposal Network under the influence of external interference factors,a verification learning method for single-target pedestrian tracking is proposed.Firstly,Siamese Region Proposal Network is used to extract features and generate multiple candidate boxes,and the best target location is selected according to the ranking of candidate box scores and stored temporarily.Secondly,whether to trigger the correction mechanism is determined by the verified learning tracking results.Finally,the corrected results are fed back to the tracker again,and location as well as size of the target are updated to complete continuous tracking of the target.Experiments on VOT2018 dataset show that the expected average overlap of our method is 0.331.The correction mechanism as well as the feedback mechanism ensure the generalization ability and robustness of the tracker to a certain extent.In terms of multi-target pedestrian tracking,aiming at the difficulty of frequent pedestrian interaction and occlusion in the video,a multi-target pedestrian tracking method combining the appearance characteristic is proposed.Firstly,a residual network is designed,which is integrated with SE module.It is trained by person re-identification dataset,and four different loss functions are used to optimize the model.Secondly,pedestrians are detected by YOLOv3 detector,then the detection results are sent into the trained network to extract features.The state of pedestrian is updated by combining Kalman filter to predict the motion estimation.Finally,the association cost matrix is constructed by the detection and tracking results,and the Hungarian algorithm is used for matching to achieve multi-target pedestrian tracking.Experiments on MOT16 dataset show that the MOTA and MOTP can achieve 53.7% and79.2% respectively.In terms of software implementation,PyQt5 is used as interface design language and Pycharm is used as development platform.Also,pytorch and tensorflow are the deep learning framework.We program and implement the video-based pedestrian tracking software.Moreover,data collected in the real scene is used for verification and test.The functions of image-based pedestrian detection,video-based pedestrian detection,single-target pedestrian tracking and multi-target pedestrian tracking are implemented.