Online Object Tracking Based On Joint Modeling With Appearance Model And Motion Estimation

Object tracking is one of the most important research topics in computer vision and pattern recognition,and due to lack of prior knowledge about the target and scene for online object tracking,and the target appearance change is unpredictable during the tracking process,which makes the online object tracking more difficult when compared with offline tracking.This thesis combines the attributes of general appearance model and accurate motion estimation,it mainly investigates online object tracking based on joint modeling with appearance model and motion estimation.By constructing the robust appearance model and flexible motion constraints,this thesis improves the adaptiveness and accuracy of the tracking model,and achieves accurate and robust real-time tracking in multiple complex scenes finally.For the irregular shape of the non-rigid object,which is hard to accurate represented with traditional bounding box method,this thesis investigate more general non-rigid object tracking with irregular image region.By incorporating the motion estimation into online object segmentation,more accurate object representation is obtained,which effectively solves the key problem of the arbitrary object tracking,and obtains good tracking results.However,when the color distribution of the target is not uniform or the background is very similar to the target,it is hard to segment accurate target region,and the performance of the proposed method would be reduced.Therefore,this thesis aims to improve the generality of the tracking model.After investigate the sample labels distribution of the online classifier,the relationship of the labeled samples and unlabeled samples is used to improve the performance of the online classifier.This thesis utilizes the spatial distribution of the sample labels and motion estimation between two consecutive frames,a reformative transductive learning method based on spatial constraints and motion is proposed for online tracking,which effectively reduces the computation cost of the clustering procedure of the traditional transductive learning,and improves the effectiveness of the classifier learning.The proposed optional strategy for target prediction and budget buffer based method for online model update make the learned classifier more robust,and overcome the inaccurate object prediction problem when the target is similar to the background.The proposed tracker performances good in various complex scenes and the effectiveness of the online learning has improved,however the accurate and real-time online tracking still need to be investigated.In order to develop an accurate and real-time online tracker,this thesis mainly focuses on improving the real-time performance of the tracking algorithm and guaranteeing its accuracy at the same time.For fast object prediction and accurate object localization,a high order regularization tracking fusion framework based on appearance model and motion estimation is presented,which combines the advantages of the robust appearance model and accurate motion estimation.By constructing a circulant structure kernel matrix as the appearance regularizer for fast object prediction,and using the motion regularizer for adaptive target scale estimation and fast movement.Finally,more accurate and robust tracking model with the high order regularization fusion is achieved.Compared with many state-of-art trackers on benchmark dataset,which demonstrates the advancement and effectiveness of the proposed tracking algorithm.