Object tracking is a hot research topic in computer vision.It has been applied in human computer interface,intelligent monitoring,behavior recognition,military guidance and so on.In real world,object tracking usually suffers from shape deformation,pose change,occlusion,illumination change,etc.So it is still a challenging task to design a robust object tracking algorithm.There are two issues need to be addressed in object tracking:(1)how to model the object appearance feature,and(2)how to choose the tracked object from the candidates samples.In order to handle these two problems,we focuses on the research of object tracking via discriminative metric learning.The main contributions of this thesis can be summarized as follows:1.This paper presents an improved bias disctiminative component analysis for visual tracking.To resolve the problem that the single feature can not adapt to the changing scenes,we construct a multi-feature appearance model,which consist of color,texture and orientation information.An improved bias discriminative component analysis classifier(BDCA)is utilized to learn a distance metric matrix,which projects the original feature space into a new transformed space.The tracked object can be located in the new metric space by matching the candidate image regions with templates in the library.Moreover,to reduce the search space,mean shift algorithm is also used to predicate the rough location in current frame,therefore the candidate image regions are collected around the rough location.The experimental results show that the proposed method is able to deal with the object appearance change,and robustly track the target object under complex scenes,such as posed change,rotation and occlusion,etc.2.We propose an online discriminative metric learning based hybrid appearance model,which combined the discriminative global and generative local appearance models.A compact global object representation is developed by extracting the low-frequency coefficients of the object color and texture based on the two dimensions discrete cosine transform(2D-DCT).Then,based on the global appearance representation,we learn a discriminative metric classifier in an online fashion to differentiate the target object from the background,which is very important to robustly indicate the appearance changes.To overcome the shortcomings of discriminative model which can not distinguish the local spatial structure of the object well,we bulid a generative local model which encodes the scale invariant feature(SIFT)and the spatial geometric information.Furthermore,different mechanisms are used to update the global and local templates to capture appearance changes.Finally,these two models are incorporated into Bayesian inference framework to effectively improve the robustness of the object tracking.3.To avoid the drifting problem in discriminative object tracking,we propose a novel object tracking method using structural sparse representation based semi-supervised learning and edge detection.Based on the clustering dictionary learned from the local image patches,the object structured sparse representation is constructed by extracting sparse code features on different layers.This structured sparse representation exploits the global and local sparse representation information,and also takes the spatial relationship between local patches into account.To utilize unlabeled samples information,a semi-supervised learning method(Laplacian Regularized Least Squares)is introduced to train a classifier to measure candidates.In addition,an auxiliary positive sample set is maintained to improve the performance of the classifier.We subsequently adopt an edge detection proposal to tune the classification results and alleviate the drifting problem.Finally,we update the classifier and edge feature template to handle the object appearance variation.Compared with several new tracking algorithms,the proposed method is able to achieve better tracking results. |