A neural network approach to human posture classification and fall detection using RGB-D camera

In this paper, we describe a human posture classification and a falling detector module suitable for smart homes and assisted living solutions. The system uses a neural network that processes the human joints produced by a skeleton tracker using the depth streams of an RGB-D sensor. The neural network is able to recognize standing, sitting and lying postures. Using only the depth maps from the sensor, the system can work in poor light conditions and guarantees the privacy of the person. The neural network is trained with a dataset produced with the Kinect tracker, but it is also tested with a different human tracker (NiTE). In particular, the aim of this work is to analyse the behaviour of the neural network even when the position of the extracted joints is not reliable and the provided skeleton is confused. Real-time tests have been carried out covering the whole operative range of the sensor (up to 3.5 m). Experimental results have shown an overall accuracy of 98.3% using the NiTE tracker for the falling tests, with the worst accuracy of 97.5%.