Finger-Worn Device Based Hand Gesture Recognition Using Long Short-Term Memory

Body-worn device based human activity recognition has far reaching influence on our life. Gesture based input, as a natural human machine interaction, will play an important role in many application fields. In this work, we propose a recognition system of hand gesture as a natural command input. The system uses a ring-shape device to capture hand movement and employ Long Short-term Memory networks to identify the gestures. We define 24 kinds of natural hand gestures and design five confusion-increasing gesture sets for system test. We analyze the confusion of the gesture sets by DTW-based similarity measure to quantify whether the elements in a set is easier to be classified than other sets. Evaluation is given on a 1680 samples dataset which analyzes the influence of system parameters on recognition accuracy and discusses the system performance under the gesture sets. The result shows the sensitivity of the finger-worn device capturing hand gesture and the stability of recognition method by comparing with other methods.