Image feature based video object description and tracking

Abstract

This paper presents a new object description method based on SIFT image features. A set of images containing targeted objects are trained. Towards extraction of image features, each image is preprocessed by applying different perspective planar transformations, and a set of points, which are robust with respect to geometrical deformations, is obtained. These transformations are chosen in a manner to preserve the perceptional identities of the principal objects existing in the transformed images. The main contribution of this study consists of comparing the trained images with the transformed images and gathering a set of the most stable points which are representing the principal objects of the trained images. These stable points derived by the set of the trained images, are then used as a robust description and tracking of the objects in motion. In order to improve reliability of the presented method, an algorithm is proposed to correct the mismatches which occur at point matching stage. The results of the studied method are compared with classical SIFT matching. Better results illustrate the effectiveness and the robustness of the SIFT based object description.