Class-driven Bayesian background modelling for video object segmentation

A background subtraction video segmentation algorithm that works by modelling the different appearances of a pixel in a set of independent layers is proposed. The main contribution of this work with respect to the existing approaches is the use of an a priori classification scheme that classifies the pixel before updating the background model. This scheme isolates the pixel instances that belong to the foreground, hence avoiding their influence in the model updating and discrimination processes of the subsequent frames. The presented results demonstrate the successful performance of the algorithm in the presence of highly dynamic backgrounds, foreground-background similarity, hot starts and abrupt illumination changes.