A fast and robust real-time surveillance video stitching method

Abstract

Abstract Real-time video stitching can build a wider field of view for surveillance, which faces a compromise between stitching speed and visual quality. A fast and robust real-time surveillance video stitching method is proposed to deal with the ghosting effect caused by moving objects and misalignments caused by background change or slight camera shift through automatic updating. By stitching key frames, parameters such as pix mapping table, stitching seams and blending weights are calculated, and most of subsequent frames are directly blended with CUDA acceleration based on the pre-calculated stitching parameters. Fast and effective algorithms are designed to detect the change of stitching seam and background during the whole stitching process, which determines whether to update the stitching seam or recalculate stitching parameters. Experiments show that this method can robustly and automatically solve the ghosting and misalignments to improve visual quality and achieve satisfactory real-time performance.