Vision-based traffic accident detection using matrix approximation
Siyu Xia, Xiong Jian, Ying Liu, Gang Li · 2015
Vision-based traffic accident detection is a significant task in traffic video surveillance. In this paper, we propose a fast and effective approach to automatically detect traffic accident in a video. The key idea is to utilize low-rank matrix approximation based method. A critical observation that traffic accidents usually occur on road area and occupy a small part of image enlightens us the following research. Each frame is first divided into non-overlapping blocks associated with different weights based on the average velocity magnitude of blocks in the training time. The motion matrix of a video segmentation is then extracted. After using low-rank matrix approximation to associate normal traffic scenes with a set of motion subspaces, we identify traffic accident at the moment of the increase of approximation error. Experimental results on several surveillance videos demonstrate the effectiveness of our proposed method.