Multi-cue collaborative kernel tracking with cross ratio invariant constraint

Lili Ma, Jian Cheng, Hanqing Lu

Abstract: In this paper, a novel multi-cue collaborative kernel tracking algorithm is proposed. A new constraint based on the property of cross ratio invariant enables tracking of objects insensitive to complex motions, including scale changes, rotation and especially views changes, without labeling and training. Meanwhile, invariant moments are introduced into the kernel based tracking method as the shape representation. The integration of shape and color information makes tracking more robust, and avoids the kernels drifting when color information is not sufficient. Experiments show our method is robust to arbitrary motions of articulated objects and other rigid objects in complex environment.