Multi-factor analysis of texture and color-texture features for robust hand detection in non-ideal conditions

Present paper explores gray-level texture features and their extension to color spaces to check their robustness and feasibility in hand detection system. A bare hand detection is affected by uneven illumination, skin tone variation, affine distortions, position variation, complex background, etc., which also effects the performance of features as well as learning ability of classifiers. Despite all, making gestures using bare hand comes naturally to humans as compared to any input devices such as red-markers, data-glove, and 3D cameras etc. In this system, three texture features, namely, gray-level histogram, R-HoG, Gabor feature and three color-texture features, namely, color based HSV histogram, color R-HoG, and color Gabor, are explored using Naive Bayes (Probabilistic view), Euclidean and Chebyshev distance (Proximity view) and Gentle, Modest and Real AdaBoost (Boosting algorithm). The features are extracted from three images resolutions, 50×50 pixels, 100×100 pixels and 200×200 pixels, to analyze the effect of resolution on the features. Database are self-collected and validated using Cohen's Kappa test. Analytic hierarchy process (AHP) based pairwise comparison among features and classifiers is performed to rank them based on performance and weightage of their performance w.r.t each other. Experimental results suggest that performance of all color-texture features are better than their respective texture features. However, the most robust combination is observed to be of color-HoG feature for Modest AdaBoost classifier.