Publication

Abstract : Efficient background modelling has always been an active area of research due to its immense importance as a preliminary step in various machine-vision applications. Several techniques have been proposed to date that strive to achieve higher accuracy without compromising on computational and hardware demands. One of such techniques, Visual Background Extractor (Vibe), has set benchmarks due to its fewer memory requirements and good results. However, it suffers from high false positives due to its slower, selective and random update policy. This paper proposes a novel sample-consensus pixel-based technique for efficient foreground segmentation complemented with faster ghost suppression. This is achieved by employing segmentation masks exploiting both static and dynamic properties of pixels depicting likeliness towards absorption into the foregrounds. Dynamic characteristics of the proposed approach handle `object present in the first frame' problem while static characteristics handle improper illumination and shadows in videos in lesser time. It aims not only at suppressing ghosts in the foreground mask but also allows their absorption by updating the background model with such regions. It also proposes a unique spatio-temporal model initialization technique for handling continuous noise. The proposed approach proved to produce outstanding results when compared with 9 traditional and 13 state-of-the-art algorithms.