Human Falling motion detecting system

Abstract

Fall accidents has been a problem for a long time. In this paper, I proposed a fall detection approach by analyzing the skeletal joints data using a machine learning tool Visual Gesture builder. Compared with past research which focuses on discrete detection like posture recognition or height determination, my approach evaluates a fall action as a sequence of frame action which means a decrease of false positive rate. A machine learning Visual Gesture Builder is used in this research. The tool allows the usage of a discrete classifier(trained by adaboost algorithm) to detect discrete gesture and a continuous classifier(trained by random forest regression) to detect fall progress. A fall detection compensation algorithm which track the confidence of head joint and spinemid joint has also been applied to deal with occluded fall. A Windows presentation foundation application is created for the fall detection algorithm and a sms will be sent if a fall detected. The system achieved an overall accuracy of 89%.