Abstract

In this paper, we propose a novel and robust fall detection system by using a one class support vector machine based on video information. Video features, including the differences of centroid position and orientation of a voxel person over a time interval are extracted from multiple cameras. A one class support vector machine (OCSVM) is used to distinguish falls from other activities, such as walking, sitting, standing, bending or lying. Unlike the conventional OCSVM which only uses the target samples corresponding to falls for training, some non-fall samples are also used to train an enhanced OCSVM with a more accurate decision boundary. From real video sequences, the success of the method is confirmed, that is, by adding a certain number of negative samples, both high true positive detection rate and low false positive detection rate can be obtained.