Efficient Foreground Segmentation for a Pan-tilt Camera Based on GMM over Walsh Spectrums

This paper describes foreground segmentation from dynamic backgrounds by using pan-tilt/translation HDTV camera. That system is designed to be used for outdoor surveillance without using mosaiced backgrounds, which sometimes cause ghost foregrounds. The algorithm is based on GMM over Walsh spectrums on many blocked images. Local spectrum based foreground segmentation makes the proposed approach robust to sub pixel motions and even non pixel level calibration. Due to insensitive sub-pixel motion nature of the algorithm, the phase only correlation (POC) between subsequent frames is used for camera movement. In order not to use mosaiced background, the background model of newly appeared area is quickly constructed by accelerating adaptation speed and spatial prediction. The experimental results show stable operations.