Detecting falls and poses in image silhouettes

Abstract

About one third of all people aged 65 and above will accidentally fall during one year. A fall can have severe consequences,such as fractures, and a fallen person might need assistance at getting up again. A lot of research has been dedicated into the development of automatic fall detection methods during the recent years. These automatic methods are created to detect falls so an alarm can be raised and help can come. In this thesis, a part of a fall detection system for a household robot aimed at helping the elderly is developed. The system is able to classify human pose from a silhouette in an image. By associating the pose “lying down” with a fallen person, the system can be used for fall detection. The algorithm is based on an image analysis feature called shape contexts. These shape contexts describe distributions of edge points by binning them into polar histograms. Altough the dataset used for training contains falls in many difficult angles, the algorithm classifies falls correctly for 97 % of a set of unseen images.