Subspace Bootstrapping and Learning for Background Subtraction

Abstract

A new background subtraction algorithm is proposed based on using a subspace model. The key components of the algorithm include a novel method for initializing the subspace and a robust update framework for continuously learning and improving the model. Unlike traditional subspace techniques the proposed approach does not require supervised or lengthy training data upfront, but instead is bootstrapped using a single background frame and exploiting spatial information in place of temporal data to generate pixel statistics for the model. The update framework allows for intelligently updating the model and re-initialization if required as determined by the algorithm. Experimental results indicate that the proposed subspace algorithm out performed traditional subspace approaches and was comparable to and sometimes better than leading standard pixel-based techniques on several standard background subtraction data sets.