Real-time record sensitive background classifier (RSBC)

Abstract

In computer vision, one of the ways to identify moving objects or foregrounds from a complex video scene is background subtraction. Two problems that usually arise in background subtraction are that of dynamic background and ghost effect. In this article, an efficient, adaptive, real-time, novel background subtraction method is proposed to handle these two problems by utilizing the classification history of previous frames to detect changes in a video scene. The proposed method, coined as Record Sensitive Background Classifier (RSBC) is an amalgamation of two models, the Adaptive Background Model (ABM) and the Neighbourhood Background Model (NBM), which make use of the foreground and background classification information of the previous frames. The novel design of the ABM considers the temporal history of transitions between foreground and background of a given pixel, while the novelty in NBM lies in considering the binary classification results of the local neighbourhood in the preceding frame. Several state-of-the-art methods are compared with the proposed RSBC method, over which the proposed method shows significant improvement. Because of its good performance at fast processing speeds the proposed method can also be used as a preliminary prioritization step for slower but high accuracy methods in an object tracking system.