Gesture recognition with inertial sensors and optimized DTW prototypes

In this work our approach for human gesture recognition with inertial sensors is presented. The proposed method utilizes a dynamic time warping (DTW) algorithm for online time series recognition. Our DTW implementation is able to deal with gesture signals varying in amplitude and to resolve ambiguities in the recognition result when DTW is used for multiclass classification. In order to find representative prototypes, an optimization method is proposed basing on an evolution strategy. By means of this prototype optimization method, we search for prototypes that yield a good class separation. Furthermore, the optimization is able to cope with approximately labelled training data, since optimal prototypes can be found in subseries of class templates. The approach has been tested with a dataset of hand gesture sequences, which have been recorded with accelerometers. Our evaluation shows that our method achieves high precision rates and good recall rates in user-dependent online gesture recognition.