Content-based video database retrieval through robust corner tracking

Abstract

A content-based video retrieval system based on extracting corners from frames and tracking them through video databases is proposed. To extract corners from each frame of video sequence in pre-processing stage, proposed multi-scale corner detector is applied. As a user interface, a proposed fast active contour model has been used to specify one object of interest as a query in one of the frames of each video shot. This frame which is in the selected-frame are extracted and tracked forwardly and backwardly through the whole of that shot using proposed multiple-match tracker. The multi-match tracker, which does not make any important assumptions or use any motion models, can retrieve the query in any video sequence even when there is unconstrained and non-smooth motion. By tracking the corners or query object forwardly and backwardly, the positions of similar objects in each video frame are determined. Two methods are considered for demonstrating the query and its similar objects to the user. Experiments have been carried out on a wide range of real video databases. All the results confirm that among the existing techniques, proposed method which exploits image corners is believed to be more generally applicable.