Research On Fusion-based Visual Object Tracking Algorithms

Visual object tracking is an important research aspect in computer vision and hasbecome a popular issue in video surveillance systems,3D scene reconstruction andhuman-computer interaction, etc. Thus the research on visual object trackingalgorithm is of great significance. The realization of robust object tracking algorithmsis confronted with several difficulties: pose and illumination variation, occlusion andbackground clutter, etc. To resolve the problems, a large number of effective trackingmethods are proposed in the past decades. Particularly, the fusion-based trackingstrategy outstands and gets widespread concern for its good performance. The subjectof our paper is just about the research on fusion-based visual tracking algorithm.Firstly, we propose an adaptive fusion based target tracking method, in which thefusion algorithm is derived from tracking discrimination and stability. Discriminationand stability have intuitive physical meanings: discrimination measures the differencebetween target and background, while stability measures the deviation degree fromtrue target center to tracking result. Algorithmically, discrimination and stability aremodeled separately first. Then, we introduce them into an adaptive fusion framework,and further formulate an object function. Finally, this object function is optimized toobtain the adaptive fusion weights. Comparative experiments on different kinds ofvideos show that our algorithm holds higher tracking precision and stability.Moreover, a kernel sparse representation (KSR) based algorithm withmulti-kernel fusion for visual tracking is presented too. Our method reconsiders thesparse representation (SR) in kernel space. The introduction of kernel scheme into SRenables us to implement multi-feature fusion strategy via multi-kernel integration.Besides, we propose a kernel coordinate descent algorithm to optimize the kernelsparse representation expediently. We implement the KSR based tracking algorithmunder particle filter framework and comparative experiments on challenging scenesdemonstrate this tracking algorithm is quite robust and achieves higher accuracy.