Fall Recognition Approach Based on Human Skeleton Information

In this paper, we present a fast approach for fall recognition. This approach according to human-body skeleton information which was obtained from Kinect sensor. First, following the falls defined by FICSIT, head and center joints, and their relative distance are considered as feature to describe the behavior. Second, applying the slide-window method and threshold for behavior action stage, motion feature vector was extracted. In the end, falls were trained and recognized by Optimal Levenberg-Marquardt BP classification algorithm. Experimental results show that the approach is high-efficiency, and rate of accuracy reaches 90%.