A Real-Time and Continuous Fall Detection Based on Skeleton Sequence

One of the leading causes of death and injury in people, especially for the elderly, is falls. Nowadays, more and more elderly people lives alone especially in Covid-19 pandemic period, therefore, it is crucial to detect falls. This paper presents a vision-based method for fall detection. Compared to non-vision-based methods, vision-based approach could not only detect falls but also provide context information that could be useful for emergency services. While a number of vision-based methods focus on fall/non-fall classification with the assumption that videos are segmented into clips. This paper introduces a real-time and continuous fall detection from video sequence. To this end, person detection and tracking are performed to determine the bounding boxes of person. Then, skeleton sequence is extracted from these detected bounding boxes. Finally, a sliding window technique is incorporated with a graph convolution network for detecting fall in continuous manner. To increase the accuracy of fall detection, a post-processing step is applied on the recognition results. The proposed method is evaluated on three datasets of fall detection: two benchmark datasets FDD and URFD and the self-built dataset named CFD. The results achieved an accuracy rate of over 99% on the FDD and URFD and over 97.14% on both view of CFD dataset at 21 fps.