Robust contour-based object tracking integrating color and edge likelihoods

摘要

We present in this paper a novel object tracking system based on 3D contour models. For this purpose, we integrate two complimentary likelihoods, denned on local color statistics and intensity edges, into a common nonlinear estimation problem. The proposed method improves robustness and adaptiv-ity with respect to challenging background and light conditions, and can be extended to multiple calibrated cameras. In order to achieve real-time capabilities for complex models, we also integrate in this framework a GPU-accelerated contour sampler, which quickly selects feature points and deals with generic shapes including polyhedral, non-convex as well as smooth surfaces, represented by polygonal meshes.