Inertial Gesture Recognition with BLSTM-RNN Gr´ egoire Lefebvre, Samuel Berlemont, Franck Mamalet, Christophe Garcia Abstract This chapter presents a new robust method for inertial MEM (MicroElec- troMechanical systems) based 3D gesture recognition. The linear acceleration and the angular velocity, respectively provided by the accelerometer and the gyrome- ter, are sampled in time resulting in 6D values at each timestep, which are used as inputs for our gesture recognition system. We propose to build a system based on Bidirectional Long Short-Term Memory Recurrent Neural Networks (BLSTM- RNN) for gesture classification from raw MEM data. We compare this system to a statistical method based on HMM (Hidden Markov Model), to a geometric ap- proach using DTW (Dynamic Time Warping), and to a specific classifier FDSVM (Frame-based Descriptor and multi-class Support Vector Machine) using filtered and denoised MEM data. Experimental results, on a dataset composed of 22 individ- uals producing 14 gestures, show that the proposed approach outperforms classical methods with an average classification rate of 95.57% and a standard deviation of 0.50 for 616 test gestures. Furthermore, these experiments underline that combining accelerometer and gyrometer data gives better results than using a single inertial description. Gr´ egoire Lefebvre Orange Labs, R&D, Meylan, France, e-mail: gregoire.lefebvre@orange.com Samuel Berlemont Orange Labs, R&D, Meylan, France, e-mail: samuel.berlemont@orange.com Franck Mamalet Orange Labs, R&D, Rennes, France, e-mail: franck.mamalet@orange.com Christophe Garcia LIRIS, UMR 5205 CNRS, INSA-Lyon, F-69621, France, e-mail: christophe.garcia@liris.cnrs.fr 1