Automatic video-object segmentation employing multi-sprites with constrained delay

摘要

This paper proposes an automatic video-object segmentation system for consumer media, based on the background subtraction technique. In order to allow for camera motion, the input images are aligned to a large panoramic background-sprite image. A multi-sprite technique is applied to minimize the size of the synthesized sprite images. In contrast to previous algorithms, which required multiple passes over the input data, the proposed algorithm enables online processing with a fixed processing delay. This is important for implementation in consumer devices which have to run in real-time with constrained resources. We provide example results illustrating that a real-time segmentation on memory-constrained hardware is feasible.