Robust Homography-Based Trajectory Transformation for Multi-Camera Scene Analysis

In this paper we present a robust image registration algorithm for the estimation of extended object trajectories in a distributed camera setting. The registration algorithm is based on a statistical homography estimation that accounts for errors related to the estimation of the homography matrix. Unlike traditional approaches based on singular value decomposition (SVD), we derive a planar homography estimation from the renormalization technique. We demonstrate the proposed algorithm with the generation of the mosaic of the observed scene as well as with the registration of the spatial locations of moving objects (trajectories) from multiple cameras. Finally, we compare the transformed trajectories with those obtained with the homography estimated by SVD and least mean square (LMS) methods and discuss the improvement in terms of spatial accuracy using objective evaluation metrics on standard test sequences.