A vision-based posture recognition system is proposed utilizing Optimal DAGSVM (Directed Acyclic Graph Support Vector Machine) classifier to achieve natural and reliable human-robot interactions. Coarse-to-fine feature detection scheme extracts skin-colored candidate regions

followed by face and hand verifications with Gabor filtered eye features and wavelet-moments of hand edge respectively. Statistical invariant moments and relative coordinates of face and hand regions are calculated as pattern feature vectors.A set of binary SVM classifiers are combined using Decision Directed Acyclic Graph with optimal structure to construct a more accurate multi-class DAGSVM classifier. Experimental result validates the reliable performance of the approach

where a natural and friendly interaction is achieved with a service robot.