Inferential Motion Reconstruction of Fall Accident Based on LSTM Neural Network

Abstract

With the rapid development of wearable technology, wearable medical devices have gradually garnered a significant amount of research interest. Motion reconstruction can accurately reproduce the posture of the user at the time of the accident, which provides medical personnel with necessary reference information. However, because of the vast range of human body activities, motion reconstruction needs high-frequency sampling data to avoid the occurrence of errors. Moreover, the fact that movements resulting from an accident can be irregular, the difficulties arising from unexpected training samples. This study attempts to establish a real-time human body inferential motion reconstruction system on fall accident. The data of human motion is recorded by using tri-axis accelerometers and tri-axis gyroscopes. The angles and tracks of the human limbs computed, and the next action occurrence point deduced using long short-term memory. Then the postural trajectory is corrected using feedback inference of gravity data from the end of a fall accident. Through the correction mechanism of bidirectional feedback, the error diffusion caused can reduce efficiency. In this study, using a parameter adjustment strategy under data sampling rate of 0.01 s, the average normal-m reconstruction rate, as well as the fall-motion reconstruction rate, can be determined. The overall posture is reproduced through the 3D video to ambulance personnel as a reference.

https://ift.tt/2Ii59y4