Iterative motion smoothing for online video stabilization in wide-area surveillance

Abstract

An online solution for video stabilization in wide-area surveillance is crucial. Unfortunately, most online methods either suffer from the error accumulation problem (drift) or produce a non-smooth camera path (jitter). We present an online approach that mitigates these drawbacks as follows. First, it detects features from video frames. Then, descriptors are extracted around the detected features to match features between frames. For motion estimation between frames, the similarity transformation is estimated. Finally, motion smoothing or compensation is performed in an iterative way that is immune to error accumulation. Our experimental results on wide area motion imagery (WAMI) show that our approach produces a smooth camera path and is robust against error accumulation. Moreover, our results are comparable to those of state-of-the-art offline methods.