Background Recovery in Railroad Crossing Videos via Incremental Low-Rank Matrix Decomposition

Inspired by the recent success of low-rank matrix recovery, we propose a novel incremental learning algorithm based on low-rank matrix decomposition. Our proposed algorithm can be applied for solving background removal problems from static yet time-varying scenes. And, in this paper, we particularly consider background modeling for railroad crossing videos. The success of an adaptive background modeling/removal approach like ours will allow users to automatically perform foreground (or intruder) detection on such scenes, which would prevent possible vehicle-train collisions and thus significantly reduce the fatality or injury rates. The challenges of background modeling in railroad crossing videos not only involve environmental variations like lighting or weather changes, headlight reflection on rails caused by nearby vehicles and foreground objects with very different velocities (e.g., vehicle, bikes, or pedestrian) also make background removal of such real-world scenes extremely difficult. We will verify that our proposed algorithm exhibits sufficient effectiveness and robustness in solving this problem. Our experiments on real-world video data would confirm that, while our approach outperforms baseline or state-of-the-art background modeling methods, our computation cost is significantly lower than that of standard low-rank based algorithm.