Human Fall Detection using Depth Videos

Abstract

The proposed fall detection approach is aimed at building a support system for old age people living alone in their homes. In this work, a method is proposed based on depth videos. A Region of interest (ROI) is detected by subtracting background from extracted frames along with preprocessing such as filtration, binarization and connected component analysis. The threshold is calculated by contemplating ROI points. Comparing ROI in each frame with calculated threshold, fall is detected. To scrutinize fall detection approach, videos of fall and no-fall activities from UR fall dataset and SDU fall dataset are processed. The results show 100% accuracy for fall activities and 82.50% for no-fall activities with UR fall dataset. Also SDU fall dataset shows 100% accuracy for fall and 80% for no-fall.