Abstract
Fall events are important health issues in elderly living environments like homes. Hence, a confident and real-time video surveillance device that pays attention could better their everyday lives. We proposed an optical flow feedback convolutional neural network according to the video stream in a home environment. Our proposed model uses rule-based filters before an input convolutional layer and the recorded optical flow for supervising the optical flow of variation. Detecting human posture is a key factor while fall events are like a falling posture. By sequencing frames of action it is possible to recognize a fall. Our system can clearly detect the normal lying posture and lying after falling. Our proposed method can efficiently detect action motion and recognize the action posture. We compared the performance with other standard benchmark datasets and deployed our model to simulate a real home situation, and the correct ratio achieved 82.7% and 98% separately.
Funding Information
  • Ministry of Science and Technology, Taiwan, China (106-2511-S-019-003, 105-2511-S-218 -003 -MY2)

This publication has 87 references indexed in Scilit: