Research Of Multiple Pedestrian Tracking Algorithm Based On Multiple Features

In modern world,the development of video surveillance has provided a guarantee for people’s travel safety to a large extent.However,with the increasing amount of data and the increasingly complex scenarios,the disadvantages of relying on manual surveillance have become increasingly prominent and the intelligent surveillance systems have also emerged.The most important and difficult part of a high-quality video surveillance is object detection and object tracking.How to accurately detect and track pedestrians or vehicles in the videos has become a hot research topic.Based on this,the paper designs a multiple pedestrian tracking algorithm,the algorithm integrated multiple features include appearance,position and motion,then realized one-step tracking with detection and data association,it also provided a research direction for this field.The paper designed a feature extraction algorithm,which is different from the algorithm that extracts features directly on the feature map,the algorithm in this paper cascades the features extracted from each layer,thereby improving its characterization ability.and based on the FoveaBox detection algorithm,a target tracking network that integrate the detection branch and the data association branch is built using the mmdetection framework.Unlike most of tracking algorithm using the public detectors and put their focus on the data association,the paper detects pedestrians using private detector.In the data association part,in addition to the appearance features extracted by the association branch,the kalman filter was also used to fuse the object’s location and motion features to match the trajectories and the pedestrians.For unmatched ones,the IoU overlap matching is utilized to maximize the tracking performance.In experiments,this paper selected three datasets of MOT 16,MOT 17 and MOT20 in MOTChallenge for training and testing,two pedestrian re-identification datasets CUHK-SYSU and PRW are used to assist training.In the detection branch,in order to deal with the missed detection of small targets in crowded scenes,this paper used FPN pyramid for multi-scale prediction.In the training phase,the CrowdHuman dataset that focuses on dense pedestrian scenes was used to pre-train the detection branch,so that the model can better cope with crowded scenes.The paper implemented two sets of comparative experiments to determine the loss function of the associated branch,and three sets of comparative experiments to the matching strategy.After trained the model,the paper evaluates and compares with other algorithms on the MOT16,MOT17 and MOT20 test sets.The ML indicator exceeded other algorithms.Finally,the tracking results of some representative scenarios in the test set are analyzed.