FAST: Fast Adaptive Silhouette Area based Template Matching

Abstract

Template matching is a well-proven approach in the area of articulated object tracking. Matching accuracy and computation time of template matching are essential and yet often conflicting goals. In this paper, we present a novel, adaptive template matching approach based on the silhouette area of the articulated object. With our approach, the ratio between accuracy and speed simply is a modifiable parameter, and, even at high accuracy, it is still faster than a state-of-the-art approach. We approximate the silhouette area by a small set of axis-aligned rectangles. Utilizing the integral image, we can thus compare a silhouette with an input image at an arbitrary position independently of the resolution of the input image. In addition, our rectangle covering yields a very memory efficient representation of templates. Furthermore, we present a new method to build a template hierarchy optimized for our rectangular representation of template silhouettes. With the template hierarchy, the complexity of our matching method for n templates is O(log n) and independent of the input resolution. For example, a set of 3000 templates can be matched in 2.3 ms. Overall, our novel methods are an important contribution to a complete system for tracking articulated objects.