PA-Search: Predicting units adaptive motion search for surveillance video coding

Abstract

The large scale of surveillance video and the high requirement of compression in time requires a low complexity and high efficiency compression algorithm to compress surveillance video. Motion search is a very time-consuming procedure in video coding. In the recent video coding standards such as HEVC/H.265, this procedure becomes more flexible by utilizing the division structure of Coding Units (CUs) and Predicting Units (PUs). However, for surveillance videos that are often captured by fixed-view cameras, the used motion search strategy still does not make full use of their intrinsic characteristics. To address this problem, we propose a PU-Adaptive Search (PA-Search) method for surveillance videos. In PA-Search, a background model is firstly constructed for a super group of pictures and then a background-foreground representation (BFR) is derived for each frame in this group. Utilizing the BFR, PUs are classified into four categories, namely, Full Background PUs (FBPUs), Background PUs (BPUs), Foreground PUs (FPUs), and hybrid foreground-background PUs (XPUs). In PA-Search, zero motion vector (zero-MV) and non-sub-pixel search are assigned to FBPUs and an error-tolerant search algorithm is also performed to reduce the influence of PU mis-classifications; while for non-FBPUs, adaptive search range is calculated according to the PU category and its size, and a BFR-based early-termination algorithm is also used to reduce the search complexity. Moreover, an early terminate partition algorithm is adopted by Full Background CUs to further reduce the encoding time. Experimental results demonstrate the advantage of the proposed PA-Search on HEVC reference software HM-16.0. PA-Search can reduce the number of search points and the total encoding time averagely by 66.90% and 46.69% over TZ Search, while maintaining the coding efficiency.