| Résumé |
In this work, we present a realtime system for continuous gesture segmentation and recog- nition. The model is an extension of the system called Gesture Follower developed at Ircam, which is an hybrid model between Dynamic Time Warping and Hidden Markov Models. This previous model allows for a realtime temporal alignment between a template and an input gesture. Our model extends it by proposing a higher-level structure which models the switching between templates. Taking advantage of a representation as a Dynamic Bayesian Net- works, the time complexity of the inference algorithms is reduced from cubic to linear in the length of the observation sequence. We propose various segmentation methods, both offline and realtime. A quantitative evaluation of the proposed model on accelerometer sensor data provides a comparison with the Segmental Hidden Markov Model, and we discuss several sub-optimal methods for realtime segmentation. Our model reveals able to handle signal distortions due to speed variations in the execution of gestures. Finally, a musical application is outlined in a case study about the segmentation of violin bow strokes. |
