Object handoff between uncalibrated views without planar ground assumption

We present a new method of object handoff between two overlapping views in a visual surveillance system. This method requires neither camera calibration nor planar ground assumption. Our approach is composed of two phases: training a handoff table and running object handoff online. In the first phase, a handoff table which contains point correspondences between two views, is constructed from a pair of image sequences. The point correspondences are obtained by computing the conditional probability that for a point pair, given a point belonging to the foreground region at a time instant, the other also belongs to the foreground region at the same time. In the second phase, the actual object handoff is performed with the help of the table, where an optimal bipartite matching algorithm is applied to resolve matching ambiguity. The experimental results for an outdoor scene and synthetic data are shown.