Matching Deformable Features Based on Oriented Multi-scale Filter Banks

摘要

This paper presents a technique to enable deformable objects to be matched throughout video sequences based on the information provided by multi-scale Gaussian derivative filter banks. We show that this technique is robust enough for viewpoint changes, lighting changes, large motions of the matched object and small changes in rotation and scale. Unlike other well-known color-based techniques, this technique only uses the gray level values of the image. The proposed algorithm is mainly based on the definition of a particular multi-scale template model and a similarity measure for template matching. The matching approach has been tested on video sequences acquired with a conventional webcam showing a promising behavior with this kind of low-quality images.