PROJECT TITLE :

Moving Object Detection in Video via Hierarchical Modeling and Alternating Optimization

ABSTRACT:

Traditionally, video modelling experts believe that the background is the primary focus, and the foreground is created by subtracting it from the background. Hierarchical modelling and Alternating Optimization is a combined estimation problem based on the concept that foreground and background are two sides of the same coin (HMAO). In order to effectively incorporate prior knowledge about the discrepancy between background and foreground, we have developed a hierarchical extension of background and foreground models. A Markov random field is first constructed at a spatially low resolution as a pivot to assist noise-resilient refinement at higher resolutions, with the goal of better describing the class of videos with dynamic backdrop. Using the alternating direction multipliers method, we show how to incrementally improve the joint estimates of both models based on their hierarchical extensions. Our method provides a more distinct backdrop and is more resilient to noise than other competing methods, according to experimental results. For video with dynamic and complex background, HMAO provides at least comparable and typically superior results when compared to current state-of-the-art approaches.