Block-Based Global and Multiple-Reference Scheme for Surveillance Video Coding

There are often lots of periodic motions in the background of surveillance videos, such as countdown traffic lights, LED billboards and etc. The conventional motion-compensation scheme and the existing frame-based single background reference scheme cannot eliminate this kind of redundancies efficiently, especially when the cycle time exceeds the maximum GOP size. In this paper, we propose a block-based global and multiple-reference scheme to solve this problem. Firstly, the background is modeled on the basis of co-located blocks but not frames, which makes it possible to realize an adaptive block-level background updating. Secondly, multiple background blocks can be kept for one block location, which makes it suitable for modeling periodic background. Thirdly, the scheme enables global reference, which further eliminates the extensively existed redundancies among GOPs in surveillance videos. Experimental results show that the proposed scheme achieves better rate-distortion performance over the existing frame-based single background reference scheme in most cases.