Robust tracking with parts in the presence of severe occlusion.

摘要

Visual tracking of a known object in a sequence of frames can fail due to partial occlusion of the object and noise. Many algorithms used throughout the computer vision community employ robust estimators to cope with these obstacles, but achieve reliability only up to 30% or 40% occlusion, or even less in the presence of cluttered background. Clearly, many real world applications require handling higher percentages of occlusion.; We propose a new tracking algorithm based on a gradient descent search. The algorithm tracks localized parts of the object, rather than the object itself, such that not all parts are occluded at the same time. Now each part can withstand up to 30% or 40% occlusion, enabling a much higher occlusion percentage for the whole object. We also propose a novel gradient which offers amongst other advantages: a lower frame rate, increased computational efficiency, and greater robustness to noise.; Successful implementation of the proposed tracking algorithm hinges upon finding a suitable set of parts. We propose a novel approach to segmentation of the object into parts, designed to optimize robustness to occlusion for visual tracking. This segmentation differs from “standard” or “natural” segmentations based on regions of homogeneous color/texture defined by physical boundaries, as the segmentation results in parts exhibiting features desirable for tracking based on the gradient descent search.; Finally we propose a recurrent neural network to model the spatio-temporal of occlusion amongst the parts of an object. It serves to predict and isolate the occlusion of the object in a frame, diverting the occluded parts to reinforce the unoccluded parts. In all, the proposed tracking algorithm based on the parts resulting from our proposed segmentation, employing our proposed gradient, and coupled with our proposed neural network, can achieve up to 75% occlusion tolerance for certain objects, as evidenced through numerical data accumulated from extensive testing as well as real tracking sequences.