A systematic review of deep learning frameworks for moving object segmentation

Moving Object Segmentation (MOS) in machine learning and computer vision is gaining much interest among researchers in recent times. This field of dynamic scene understanding aims to understand every pixel of a video by segmenting objects in a frame sequence from beginning till end. The goal is to generate temporally consistent and accurate pixel masks for the target object in a video. Our paper follows a systematic way to represent the available state-of-the-art literature in the field of MOS. It includes a detailed analysis of semi-supervised, unsupervised, interactive, and referring MOS. Next, various deep learning-based approaches for MOS are elaborated. In Section 4, the basic architecture of MOS has been explained with a summary of the current benchmark data sets used. Further, a summary of evaluation metrics and qualitative and quantitative results forMOS are discussed. Finally, latest available results to date are provided as well as future research directions for MOS are detailed.