Joint Positions Detection for the Elderly Exercises using Backpropagation Neural Network

Porawat Visutsak, Teeradon Chaiwong, Khemmarin Jirawangkaewworrawut, Mohamed Daoudi


The inspiration of this work is based on the limitation of the elderly’s movement. The lack of monitoring while the elderly is doing the exercises may cause the injury. This system can prevent the injury of the elderly by real time detecting and alerting. The video clips of the elderly exercises used in this system were created by the sport science specialists from Thai Health Promotion Foundation (THPF). There are 2 modes of exercise: sit mode and standing mode, the elderly may prompt to see the demonstration of each exercise from the video clip on the GUI. The Kinect detects the whole-body movement, and the system will alert if the wrong posture is detected. Our method used the backpropagation neural network for training and testing models. The 16 exercises of the elderly were recorded and extracted to 8,300 frames as the training data. Our model used 25 hidden layers and 2 outputs. In the real time classification, the system yields 89.37% of accuracy. The user interface is quite simple and easy to use for the elderly.

Full Text:



  • There are currently no refbacks.