Abstract
In this paper, we present Watchful-Eye, a 3D skeleton-based system to monitor a physically disabled person using a cane as a mobility aid. Watchful-Eye detects fall occurrences using skeleton tracking with a Microsoft Kinect camera. Compared to existing systems, it has the merit of detecting various types of fall under multiple scenarios and postures, while using a small set of features extracted from Kinect captured video streams. To achieve this merit, we followed the typical machine learning process: First, we collected a rich fall detection dataset. Second, we experimentally determined the most relevant features that best-distinguish fall from non-fall frames, and the best performing classifier. As we report in this paper, the offline evaluation results show that Watchful-Eye reached an accuracy between 87.2% and 94.5% with 5.5% to 12.8% error rate depending on the used classifier. Furthermore, the online evaluation shows that it can detect falls with an accuracy between 89.47% and 100%.