Abstract

In this article a robust and real-time dynamic hand gesture recognition system meant to allow a natural interaction with a service robot, in dynamic environments, is proposed. The main novelty of the proposed approach is the use of temporal statistics about the hand's positions and velocities as basic information to recognize the gestures. The use of these features allows carrying out the final recognition using a standard Bayes classifier, instead of the traditional Hidden Markov Models. A method for simultaneous gesture segmentation and recognition, which works by finding candidate subsequences that give high scores when matched to a gesture, is proposed. The system uses boosted classifiers to detect hands, and the mean-shift algorithm for their tracking. The system's performance is validated in a digit recognition system database and in real-world video sequences.