Real-Time Accident Prediction Using Deep Learning for

Power-assist exoskeleton robots are useful for assisting activities of daily living (ADL) for physically weak persons. Such persons are also likely to have deteriorated perception ability, hence, it is important for power-assist robots to have perception-ability to ensure the safety of the user. Perception-assist can be accomplished by observing the interaction between the environment and the user, determining the possibility of accidents, such as falling, and preventing them by modifying the wearer's motion. Therefore, in order to accomplish perception-assist, it is essential to predict the possibility of accidents in real-time. In this paper, we propose a method that uses deep learning to predict the possibility of accidents based on the wearer's motion, wearer's motion intention from EMG signals, zero moment point (ZMP) and information from the surrounding environment. The effectiveness of the proposed method is evaluated by performing experiments to test accident-prediction in real-time.