Enhanced Appearance Models for Object Tracking

Abstract

This paper is concerned with improving target appearance models to realize robust object tracking. We explore the use of feature space other than the commonly used color space for object tracking. Specifically, we employ gradient information to be used separately as well as in conjunction with color information. Our target appearance model is then represented in the form of a histogram using its gradient and color feature spaces, and frame-to-frame tracking is performed using mean shift or local exhaustive search. By combining gradients with color, we build new appearance models with combined feature spaces. Based on our extensive testing of these models, we find that they can be used to track complex objects, such as full 360-degree rotating objects, appearance-changing objects, occluding objects and zooming objects.