Scale Invariant Kernel-Based Object Tracking

Traditional kernel-based object tracking methods are useful for estimating the position of objects, but inadequate for estimating the scale of objects. In this paper, we propose a novel scale invariant kernel-based object tracking (SIKBOT) algorithm for tracking fast scaling objects through image sequences. We exploit the set property of regions and propose a new method to estimate the potential of the intersection of the object and the kernel. Regarding robustness, we iteratively estimate the scale of the object by means of basic set analysis. The scale and position of objects are simultaneously estimated by mean shift procedures in parallel. The proposed SIKBOT algorithm is demonstrated by extensive experiments on challenging real-world image sequences.