Although gesture recognition has been studied for several decades, much research stays in the realm of
indoors laboratory experiments. In this thesis, we address the problem of designing a truly usable, real-
world gesture recognition system, focusing mainly on the real-time control of an outdoors robot for use by
military soldiers. The main contribution of this thesis is the development of a real-time gesture recognition
pipeline, which can be taught in a few minutes with: very sparse input (« small data »); freely user-invented
gestures; resilience to user mistakes during training; and low computation requirements. This is achieved
thanks to two key innovations: first, a stream-enabled, DTW-inspired technique to compute distances
between time series; and second, an efficient stream history analysis procedure to automatically determine
model hyperparameters without user intervention. Additionally, a custom, hardened data glove was built
and used to demonstrate successful gesture recognition and real-time robot control. We finally show this
work’s flexibility by furthermore using it beyond robot control to drive other kinds of controllable systems.
Jury members:
Dr Catherine ACHARD, Maître de Conférences, Habilitée à Diriger des Recherches, Université Pierre & Marie Curie, Paris 6
Prof. Fabien MOUTARDE, École Nationale Supérieure des Mines de Paris (ENSMP)
Prof. Éric ANQUETIL, INSA de Rennes
Prof. Pierre-François MARTEAU, Université Bretagne Sud
M. Philippe GOSSET, Responsable Industriel THALES OPTRONIQUE, Élancourt
