Codebook appearance representation for vehicle handover across disjoint-view multicameras

Abstract

Object handover as continually tracking an object across disjoint-view cameras is a necessary part of video-based monitoring systems. While having nonoverlapping cameras is a requirement for monitoring a wide area, there is no common 3D location that can be used to detect multiple views of the same object, in contrast with overlapping cameras. Appearance features play an important role for object handover in such camera networks. This paper focuses on modeling appearance features of moving vehicles by a new major color representation called codebook representation. Toward this end, in each frame, the k-means algorithm is used to cluster major colors of an object. In the subsequent frames, a set of cylinders in the RGB space called codebook keeps the track of these major colors for incremental clustering. Then, in the matching phase, a similarity measurement for comparing different codebook sets discriminates major colors of observations. In addition, a brightness transfer function is developed for mapping cylinders between two camera views. By this mapping, the model can tolerate the illumination change of environments. The method is fast enough to be used in real-time applications. Experimental results show the efficiency of the proposed methods on real datasets.