Specular Reflection Removal for Human Detection under Aquatic Environment

This paper addresses two important issues for effective human detection under aquatic environment: (i) human of interest being partly hidden by specular reflections, and (ii) non-stationary background elements. A Bayesian framework is employed using a high order Hidden Markov model with mixture transition distributions as the prior to capture similar feature information from previous frames. Foreground detection is enhanced by utilizing this model to partially or wholly recover objects hidden within specular reflections while at the same time, suppressing foreground errors attributed to dynamic backgrounds. The proposed methodology is demonstrated on an extremely challenging environment, an outdoor swimming pool, and performed robustly from day to night.