Unsupervised Background Estimation Using a Neural Integrator

Abstract

Background estimation is an important part of many computer vision applications. However, it is a challenging task due to illumination changes, camouflage, occlusion, dynamic background, rain or snow fall, and shadows. In this paper, we propose a method to predict the background of videos recorded by fixed cameras. The proposed algorithm is unsupervised and online. It takes inspiration from the processing mechanisms of neural integrator circuits in recurrently connected networks. The neural activities of three distinct integrators, each responsible for processing a color channel in L * a * b color space, are updated according to the recent changes of the scene covering both spatial and temporal aspects. The maxima of the evolving activity distributions in color space are used to predict the background color value of each pixel. Evaluation results demonstrate that the proposed method outperforms several recent competitors on the Scene Background Initialization (SBI) and LASIESTA datasets, based on mean squared error (MSE) metrics.