Abstract
A new stochastic learning algorithm for use in nonparametric pixel-level background models is presented in this paper. For the first time, we propose the use of kernel density estimation techniques in the model update step to identify outliers within the pixel-level sample collections and replace them with recently observed background pixel values. A neighborhood diffusion process that improves on recently reported scene model learning techniques is presented, wherein information sharing between similarly structured adjacent background models is encouraged to promote spatial consistency within localized image regions. We demonstrate the superiority of the proposed algorithm in comparison with the state-of-the-art visual background extraction system using the well-known percentage correct classification statistic and a new figure of merit, probability correct classification, presented here for the first time.
