Efficient and robust alignment of unsynchronized video sequences

This paper addresses the problem of aligning two unsynchronized video sequences. We present a novel approach that allows for temporal and spatial alignment of similar videos captured from independently moving cameras. The goal is to synchronize two videos of a scene such that changes between the videos can be detected automatically. This aims at applications in driver assistance or surveillance systems but we also envision applications in map building. Our approach is novel in that it adapts an efficient information retrieval framework to a computer vision problem. In addition, we extend the recent ECC image-alignment algorithm to the temporal dimension in order to improve spatial registration and enable synchro refinement. Experiments with traffic videos recorded by in-vehicle cameras dem onstrate the efficiency of the proposed method and verify its effectiveness with respect to spatio-temporal alignment accuracy.