Adaptive directional bounding box from RGB-D information for improving fall detection

Abstract

Fall detection for aging people is still a mainstream research focus for the current aging society. Tools that are simple and inexpensive but have high accuracy rates are needed. RGB-D information retrieved from a home entertainment system was used to detect falls using typical bounding boxes techniques. These techniques have limitations. This research introduced the Adaptive Directional Bounding Box that made use of a comprehensive bounding box and a dynamic state machine in a new way to detect falls. The proposed approach offered a way to track and analyze continuous data streams of the visual images to automatically predict a fall event prior to the fall state in a single-phase instead of the typical two-phases. This can significantly affect the survival or severe injury of the elderly. The proposed method can improve accuracy by 25.5% and the response time by 21.31% on average as compared to existing approaches.