A Flexible Fall Detection Framework Based on Object Detection and Motion Analysis
Dara Ros, Rui Dai
Abstract
To protect the elderly against harmful falling events, automatic fall detection solutions have been developed using wearable devices, ambient sensors, and video cameras. Vision-based fall detection is a non-invasive and low-cost solution for practical fall detection systems, as video cameras have already been deployed ubiquitously for surveillance purposes. There has been growing interest in leveraging the latest advances in artificial intelligence for image processing and computer vision to detect human falls from surveillance videos. This paper proposes a Motion-based Multi-Eye Fall Detection framework (MMEFD) based on detecting human objects and tracking objects' motion in the temporal domain. Unlike other vision-based methods that require predefined thresholds to detect a fall, the proposed framework localizes and tracks a person in videos via object detection and motion analysis over a time window with an appropriate size. As fall events may look different from different view angles, a multi-view fall dataset is used to train a classifier to detect falls. The proposed framework is flexible for different use cases as it could incorporate different object detection methods and analyze videos captured from different angles. It has produced good detection results on several other datasets that outperform two traditional fall detection methods.