DuG: Dual Speaker-based Acoustic Gesture Recognition for Humanoid Robot Control

Abstract Acoustic gesture recognition based on the Doppler effect is a natural HCI (human computer interface) mechanism for cooperation with robots. Environmental noise, nearby human motions, and user adaptability compromise the gesture recognition accuracy and weaken the efficacy of the interaction mode. To this end, we propose a novel Dual Speaker Gesture (DuG) recognition scheme in which the sensing environment consists of two speakers and a microphone. The two speakers transmit inaudible tones at different frequencies and the microphone changes in the dual-channel sound carry rich features originating from the Doppler shifts caused by hand motions, where we introduce a feature called top-bandwidth to distinguish the environment motions noise and discard. In particular, we develop a fusion framework to combine a priori empirical model with a hidden Markov model(HMM) to enhance user adaptability and improve classification accuracy. A set of 11 gestures has a mean recognition rate up to 99.36% with 10 participants. And five users participating in a driving robot experiment show that it maintains average 98% accurate gesture recognition even under a common HMM model. DuG boosts the gesture recognition performance under acoustic noise and nearby environment motion interference.