Recognition of human motion with deep reinforcement learning

Human–computer interaction (HCI) has become an important research area for improving the user experience on Internet of Things (IoT) devices. In particular, gesture recognition and daily-activity recognition have attracted the interest of numerous researchers. Human motions have been predicted by analyzing accelerometer data from which features were extracted to be classified into a specific activity. However, due to the memory limitations of IoT devices, it is hard to utilize all the raw data from an accelerometer sensor. This paper proposes a deep reinforcement learning algorithm to recognize human arm movements using a commercial wearable device, the Myo armband. Agents learn the patterns that are the acceleration data of human motion. In addition, using raw accelerometer sensor data without feature extraction could make an end-to-end structure. In order to demonstrate the performance of the proposed method, a deep neural network (DNN) and a deep reinforcement learning algorithm are compared. As a result, a deep reinforcement learning agent yielded accuracy similar to a DNN using less data, and the agent could learn time-series human motion acceleration data. Copyrights © 2018 The Institute of Electronics and Information Engineers.