An algorithm for robust motion detection in video is proposed in this work. The algorithm continuously analyses the dense pixel volume formed by the current frame and its nearest neighbours in time. By assuming continuity of motion in space and time, pixels on slanted edges in this timespace pixel volume are considered to be in motion. This is in contrast to prevailing foreground-background models used for motion detection that consider a pixel's history in aggregation. By using an efficient data reduction scheme and leveraging logical bit-parallel operations of current CPUs, real-time performance is achieved even on resource-scarce embedded devices. Video surveillance applications demand for efficient algorithms which robustly detect motion across a wide variety of conditions without the need for on-site parameter adjustments. Experiments with real-world video show robust motion detection results with the proposed method, especially under conditions normally considered difficult, such as continuously changing illumination.
展开▼
