Fall detection using history triple features

Accurate identification and timely handling of involuntary events, such as falls, plays a crucial part in effective assistive environment systems. Fall detection, in particular, is quite critical, especially in households of lonely elderly people. However, the task of visually identifying a fall is challenging as there is a variety of daily activities that can be mistakenly characterized as falls. To tackle this issue, various feature extraction methods that aim to effectively distinguish unintentional falls from other everyday activities have been proposed. In this study, we examine the capability of the History Triple Features technique, based on the Trace transform, to provide noise robust and invariant to different variations features for the spatiotemporal representation of fall occurrences. The aim is to effectively detect falls among other household-related activities that usually take place indoors. For the evaluation of the algorithm the video sequences from two realistic fall detection datasets of different nature have been used. One is constructed using a ceiling mounted depth camera and the other is constructed using an RGB camera placed on arbitrary positions in different rooms. After forming the feature vectors, we train a support vector machine using a radial basis function kernel. Results show a very good response of the algorithm achieving 100% on both datasets indicating the suitability of the technique to the specific task.