Visual adaptive tracking for monocular omnidirectional camera

This paper presents a sophisticated patch-based visual tracking algorithm using an omnidirectional camera with distortion adaptation. The omnidirectional camera is modeled using the equivalent projection theory, so that a nonlinear deformed neighbourhood can be accurately estimated in the image plane, which significantly facilitates feature coding. In order to improve the omnidirectional tracking performance, a patch-based multi-feature matching method is proposed under a probability framework. In particular, the distributions of patches covering key parts of the target are weighted adaptively according to their joint-feature response, which is able to track target robustly and filter out the outliers effectively. Extensive experiments have been conducted to verify the performance of the proposed omnidirectional tracking algorithm, which obtains promising results on challenging datasets and outperforms many state-of-the-art methods.