Research On Video-based Object Tracking Method

As an intermediate task in the field of video understanding,object tracking directly affects the performance of high-level computer vision tasks such as behavior recognition,so it has attracted much attention from academia and industry.This thesis delves into video-based object tracking methods,focusing on solving object tracking problems in several complex scenes.First,for the single object tracking task in the UAV scene,this thesis alleviates the problem of camera mutation through the camera correction module based on the ORB feature,and uses the speed prediction model based on the Bayesian method to track the object roughly in the full occlusion situation.These methods achieved the fourth place in the 2020 VisDrone Challenge,proving the effectiveness and robustness of the method.Then,for the long-term single-object tracking task in complex scenes,this thesis designs a FP-verifier based on the feature pools to reverify the results of the tracker output for the problem of unstable model update.For the problem of global searching blindly,a localization module is designed to guide the tracker to perform a global search only when necessary and within a reasonable range.The tracking framework designed in this thesis outperforms the first place in the 2020 Visual Object Tracking Challenge on the LTB50 dataset.Finally,this thesis improves the multi-object tracking algorithm,FairMOT,on the security monitoring dataset MOT20.We replace the backbone network with Swin Transformer to obtain a larger receptive field,and introduce the Re-check module into the network for the problem of insufficient discrimination of features.To make full use of the historical trajectory features to enhance the target detection ability.The method designed in this thesis has been fully controlled on the pedestrian-intensive MOT20 dataset,and the evaluation results are significantly improved compared with the baseline network FairMOT.