Web proceedings papers


Adam Botos , Dmitry Chetverikov and Peter Kovacs


The Singular Value Decomposition (SVD) can be eciently used to detect motion in videos captured by a static camera. However, the SVD is computationally demanding when a large matrix - a spatiotemporal data window typically composed of ten to thirty frames - is repeatedly processed. Recently, a running (incremental) version [1] of the SVD has been proposed and applied to motion detection. Although much faster than the direct implementation, the CPU-based running SVD is still too slow for real-time processing of VGA or larger size videos. In this paper, we present a GPU implementation of the running SVD that is suitable for robust, close to real-time motion detection in challenging full-size, standard frame rate videos. The computational performances of dif ferent hardware con gurations are compared showing a signi cant gain in processing speed due to the proposed solution.


Video processing, motion detection, SVD, GPU