A Long Term Memory Recognition Framework on Multi-Complexity Motion Gestures

Most existing researches on inertial sensor based dynamic motion gestures use deterministic or stochastic methods, however, these models generally possess short term memory so that they only memorize few time steps before and ignore the historical information deeper in time. Furthermore, researchers mainly investigate on the primary level gestures, while gestures with higher complexity are more powerful in expression. In this paper, we implement an end-to-end framework for recognition on multi-complexity dynamic motion gestures using a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN). Since the lack of available motion database, we collected three databases of motion gestures in different levels of complexity. Motion gesture signals were carefully pre-processed and sent for training without feature extraction. Results on 5-folds cross validation prove that our framework has good recognition and real-time performance on different types of gestures, and shows robustness to the invalid segments, and the time consumption of recognition keeps stable when gesture classes increase.