Dynamic Gesture Recognition Based on RF Sensor and AE-LSTM Neural Network

In this paper, we present a novel deep neural network (DNN) dedicated to dynamic gesture control based on a radio-frequency (RF) type sensors. The proposed network is based on auto-encoder (AE) and long short-term memory recurrent neural network (LSTM-RNN). The encoder for reduction and effective feature extraction is obtained by unsupervised training an AE with a large untagged RF database. Further a light LSTM network is trained use a small size of tagged datasets for dynamic gesture classification. Training the network with semi-supervised learning methods greatly reduces the requirement on the size of the tagged gestures datasets. Experimental results demonstrate that our AE-LSTM method outperforms existing networks in terms of accuracy and realizability. The proposed application is used for the human-computer interaction (HCI) of intelligent home, automatic driving and robot scenes.