Robust tracking and object classification towards automated video surveillance

Abstract

This paper addresses some of the key issues in computer vision that contribute to the technical advances and system realisation for automated visual events analysis in video surveillance applications. The objectives are to robustly segment and track multiple objects in the cluttered dynamic scene, and, if required, further classify the objects into several categories, e. g. single person, group of people or car. There are two major contributions being presented. First, an effective scheme is proposed for accurate cast shadows / highlights removal with error corrections based on conditional morphological reconstruction. Second, a temporal template-based robust tracking scheme is introduced, taking account of multiple characteristic features (velocity, shape, colour) of a 2D object appearance simultaneously in accordance with their respective variances. Extensive experiments on video sequences of variety real-world scenarios are conducted, showing very promising tracking performance, and the results on PETS2001 sequences are illustrated. © Springer-Verlag 2004.