Adaptive Pixel-Based Efficient Motion Estimation for Surveillance Video Coding

This paper introduces an innovative method termed Adaptive Pixel-based Efficient Motion Estimation (APEME) designed specifically for the lossless coding of surveillance videos. Surveillance videos, characterized by significant background proportions, pose unique challenges for coding. Existing methods often employ separate search mechanisms for background and foreground regions. However, traditional block-based motion estimation encounters two primary challenges: 1) the necessity to transmit side information such as motion vector overhead, and 2) inaccurate motion estimation at object boundaries in surveillance videos. To overcome these challenges, our proposed method adopts a pixel-by-pixel motion prediction approach. We present a five-step pixel-based motion search scheme. Initially, each pixel is classified into one of three classes: background (BG), foreground (FG), or boundary (BD). Subsequently, an adaptive target window (TW) design is applied for each class. Next, global Motion Vector (MV) candidates for respective classes and class-based adaptive search schemes are employed to expedite search convergence. Finally, adaptive prediction mode selection is employed to further enhance prediction performance. Our method demonstrates superior entropy results and a substantial reduction in computations compared to competitive pixel-based search methods across various motion surveillance videos. Experimental findings also illustrate that our proposed approach outperforms competitive methods in entropy performance for non-static regions, maintaining a computational advantage.