Efficient Video Stabilization via Partial Block Phase Correlation on Edge GPUs

Cevahir Çığla; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 2152-2161

In this paper an efficient video stabilization method is introduced that exploits phase correlation on partial blocks. The approach addresses very low computational complexity on edge GPU devices for surveillance cameras. The global motion of consecutive frames due to camera vibrations along horizontal and vertical axes are extracted through phase correlation of informative blocks among the video. The novelty of the proposed approach lies within the extraction of sub-blocks with high suitability/reliability and reliable merge of multiple block correlation maps. Pipeline implementation of detection of informative blocks and utilization of 4-8 sub-blocks during the shift estimation are the key features of enabling up to 150 fps processing for HD (1920x1080) video on Nvidia Jetson nano edge GPU device without any accuracy loss. Utilizing a generic scheme and reasoning as well as efficient GPU implementations proposed approach is highly adaptable to various video content by altering sub-block choices. Considering stabilization as a pre-process step the proposed algorithm reserves sufficient computational room for further video analyses on edge compute devices.