CDoTS: Change Detection on Time Series Background for Video Foreground Segmentation

Although many adaptive background subtraction methods have been proposed for image-based foreground detection, dynamic background in the scene, such as an electronic billboard, still causes a serious problem of false alarm. Exclusion of such area from region of interest may prevent the problem, however an issue of security hole on that area becomes another concern. A method of change detection on repetitive time series background is proposed in this paper. Our method extends an adaptive multiresolution background subtraction to allow detection of time series, which is in turn used for foreground extraction on such area. The accuracy of segmentation on static background is barely changed, while that on the area of periodic change is significantly improved.