Font Size: a A A

Research Of Visual Tracking Algorithms In Complex Environments

Posted on:2017-01-16Degree:DoctorType:Dissertation
Country:ChinaCandidate:B X WangFull Text:PDF
GTID:1108330503455255Subject:Signal and Information Processing
Abstract/Summary:PDF Full Text Request
Visual object tracking is an important yet hot research topic in the field of digital image processing, artificial intelligence and so on, and it has been widely applied in intelligent transportation system, visual surveillance, human computer interaction, aerial reconnaissance, just to name a few. While many algorithms have been proposed and there has been much progress in the tracking field, it remains a challenging problem to design a robust and efficient visual tracker due to the complex background, partial occlusion, illumination changes, motion blur, background clutter, and so on. To handle these challenges, in this thesis, we focus on how to design a robust and efficient appearance model for tracking, and some novel tracking methods have been presented.1) We propose a bidirectional cooperative sparse representation tracking model. Through exploring the traditional unidirectional sparsity tracking frameworks, the forward sparsity tracking model is to use the template set to reconstruct candidate samples, and the reverse sparsity tracking one is to project the template set into a candidate space. What the two models have in common is to compute the sparse correlation coefficient matrix of candidate sample and template set. Based on this, using L2-norm constraint item, the forward and reverse sparse correlation matrix coefficients can be uniformly convergent. In comparison to conventional unidirectional sparse tracking model, the bidirectional sparse tracking model could fully excavate the sparse mapping relation of the whole candidate sample and template set. Based on the accelerated proximal gradient fast numerical method, we derive the optimum solution(in matrix form) of bidirectional sparse tracking model. As a result, it allows the candidates or templates to be calculated in parallel, which can improve the calculation efficiency to some extent. Numerical examples show that the proposed tracking algorithm has certain priority over against the conventional unidirectional sparse tracking methods.2) We present a novel and efficient discriminative object tracker. The proposed method takes advantage of the excellent classification capability of an emergent learning technique, i.e., Extreme learning machine(ELM). Specifically, object tracking is viewed as a binary classification problem, and ELM is utilized for finding the optimal separate hyperplane between the object and its backgrounds efficiently. To exploit the relevance of object information in time and space, two constraints are introduced in ELM training:(i) Pairwise metric constraint: the target observations must have different ELM outputs from those background ones; while the various target observations among successive frames should have approximate ELM outputs.(ii) Smoothness constraint: target visual variation during a short interval is often linear, and thus, the difference between the consecutive updated ELM models should be smooth. Moreover, we have derived the online updating of ELM-based object tracking model, thereby leading to more robust tracking results. Experimental evaluations on challenging sequences demonstrate that the proposed tracker can performs well in terms of the tracking robustness and efficiency.3) A novel tracking method is presented via combining sparse representation and an emerging learning technique, namely Extreme learning machine(ELM). By taking full advantage of the above two different frameworks, we can obtain a new coarse-to-fine visual tracking model. Specifically, visual tracking task can be divided into two consecutive processes. Firstly, ELM-based discriminative tracking model is used as the coarse appearance model, which is utilized to remove most of the candidate samples related to background contents efficiently, thereby reducing the disturbances and computational cost for the following sparse representation. Secondly, by using the manifold learning constraint, a new constrained sparse representation framework is advocated as the fine appearance model, in which the object candidate can be finally determined. Here, the learning results of candidate samples on the ELM classification are directly exploited to construct the manifold learning constraint term, which further combines the ELM learning and sparse representation. Besides, the presented tracker also adopts a new feature learning framework, i.e., ELM Sparse Auto Encoder, which can provide a compact and discriminative representation of the inputs and tend to achieve more robust tracking results. Experimental evaluations on challenging sequences demonstrate that the proposed coarse-to-fine tracker performs favorably against the state-of-the-art methods.
Keywords/Search Tags:visual object tracking, sparse representation, extreme learning machine, bayesian inference, manifold learning, online updating, feature learning
PDF Full Text Request
Related items