Multisource surveillance video coding with synthetic reference frame

Due to the increasing growth of surveillance data, high-efficiency surveillance video coding schemes are demanded. However, the existing conventional coding framework has difficulties in handling rotation and zooming whilst the recently proposed multisource surveillance video coding method is inflexible and prone to be affected by pose errors. To this end, we propose to combine conventional surveillance video coding and multisource surveillance video coding into one unified framework. First, global knowledge in the form of 3D model and initial textures is employed to construct a knowledge based reference frame. Meanwhile, a temporal reference frame is also generated from the reconstructed frames in decoded picture buffer. Then, they are fused to obtain the synthetic reference frame to exploit global and local information. Finally, we add the synthetic reference frame into reference picture list and rearrange the list to optimize the overall efficiency. Experimental results demonstrate the effectiveness of the proposed method over state-of-the-art anchors. (C) 2019 Elsevier Inc. All rights reserved.