Evolving Background Subtraction for Dynamic Lighting Scenarios
Bruno Sielly Jales Costa, Madeline J. Goh · 2018
This paper presents a novel approach to background subtraction based on the concepts of typicality and eccentricity data analytics and self-evolving cloud-based classification. The proposed approach creates local (i.e. specific region of an image, up to pixel level) models of normality that can be recursively updated as new data are acquired. Each normality model can represent different normal scenarios, often significantly different from each other due to dynamic lighting (e.g. moving shadows, sunspots). Such a normality model is used to describe the background for the image stream. Each pixel or image region is then classificated as background or foreground based on its calculated eccentricity level, an aggregated measurement of the intensities of the RGB channels based on previous image samples. The proposed technique is compared with nine well-known background subtraction algorithms on a set of images of a moving vehicle interior. The results obtained are very promising, especially under challenging lighting scenarios.