Human fall detection using machine vision techniques on RGB-D images

Falling represents one of the major problems faced by elderly people. In the present research, a machine vision-based system was designed. Depth map images were captured using Microsoft Kinect® camera. They were processed for extracting features and designing the detection algorithm, apply SVM classifier, to distinguish falling pose from normal pose in 70 video samples. Furthermore, another experiment was conducted on the basis of threshold on the feature of distance to the floor, with its outputs replaced SVM responses. In the fall detection algorithm, in order to calculate speed, image features were used rather than accelerometer data. Relying on depth map images and employing Open CV library, the present research outperformed similar works where color images or such devices as accelerometers were used, attaining sensitivity and specificity of 100% and 97.5%, respectively. The use of the distance of the person's centroid to the floor efficiently contributed into better results.