A Robust Shadow Removal Technique Applying For Person Localization in a Surveillance Environment
Thuy Thi Thanh Pham, Hai An Vu, Anh Tuan Pham · 2015
In this paper, we propose a new technique for removing shadow regions based on a learning-based approach. We extract two different types of the examined shadow regions, such as chromatic-based features and physical properties. Two likelihoods or shadow-matching scores are calculated from corresponding features. However, it is different from existing techniques, we take into account a density-based score fusion scheme. A likelihood ratio of shadow per nonshadow score is calculated. Probabilities of shadow and nonshadow are estimated based on approximating distributions of the shadow-matching scores using Gaussian Mixture Models. The experimental results confirm that the proposed fusion scheme outperforms existing techniques which utilize the separated features. We then deploy the proposed technique in framework of person localization system in indoor environments. Averagely, positioning error is reduced from 44.4±33.1(cm) (without shadow removal) to 13.5±18.8(cm) (with the proposed shadow removal technique). Consequently, this work contributes an effective pre-processing step in order to deploy vision-based localization services in surveillance environments.