Unsupervised moving object segmentation using background subtraction and optimal adversarial noise sample search

Abstract

Moving Objects Segmentation (MOS) is a fundamental task in many computer vision applications such as human activity analysis, visual object tracking, content based video search, traffic monitoring, surveillance, and security. MOS becomes challenging due to abrupt illumination variations, dynamic backgrounds, camouflage and scenes with bootstrapping. To address these challenges we propose a MOS algorithm exploiting multiple adversarial regularizations including conventional as well as least squares losses. More specifically, our model is trained on scene background images with the help of cross-entropy loss, least squares adversarial loss and ℓ1 loss in image space working jointly to learn the dynamic background changes. During testing, our proposed method aims to generate test image background scenes by searching optimal noise samples using joint minimization of ℓ1 loss in image space, ℓ1 loss in feature space, and discriminator least squares loss. These loss functions force the generator to synthesize dynamic backgrounds similar to the test sequences which upon subtraction results in moving objects segmentation. Experimental evaluations on five benchmark datasets have shown excellent performance of the proposed algorithm compared to the twenty one existing state-of-the-art methods.