Video object segmentation using background model based on pixel time series clustering

摘要

Video object segmentation is an essential procedure of pedestrian detection and tracking to measure building occupancy. To tackle the difficulties of environmental light change and dynamic background in the scene of building entrance, this paper proposes a novel video object segmentation algorithm using background model based on pixel time series clustering. The background model uses specific distance metrics of pixel values in YUV space to decrease interference of environmental light change. It maintains multiple clustering centers to process dynamic background and introduces the maximum time for no match to exclude foreground objects from the background model. It uses a small storage to memorize a structured movement and builds a compact model for dynamic background. Experimental results demonstrate a good segmentation accuracy and a fast processing speed in comparison to other algorithms. Our algorithm is very suitable for scenarios such as building entrances.