Non-orthogonal Binary Expansion of Gabor Filters with Applications in Object Tracking

Gabor filter response is widely used in many computer vision applications for its effectiveness in representing local image details. The major drawback of Gabor features is the high computation cost involved in the convolution between the image and the filter bank. This paper presents a method to approximate the Gabor filters as a linear combination of Haar-like features. These features are selected from a large redundant feature pool using a generative feature selection scheme - optimized orthogonal matching pursuit (OOMP). Major advantage of this representation is that the convolution between the image and the approximated Gabor filters can be computed very efficiently using integral image trick. We applied the proposed method to object tracking, promising results are demonstrated.