Real-time Detection of Specific Events: A Case Study of Detecting Falls

Abstract

Detecting people who have fallen is a crucial problem since they may have severe injuries. In this study, we combined existing datasets to create a new dataset for generalized fall detection performance in a wild environment with diverse domains. Furthermore, we propose simple yet powerful rule-based methods for fall detection and real-time operation: the bounding box ratio and bounding box overlap. Our method was evaluated with YOLOv5 as a backbone network and achieved performance improvements by 0.126 in precision, 0.08 in recall, 0.156 in 50 mAP, and 0.11 in 95 mAP compared to our baseline, the VFP290K dataset [7]. In addition, compared to the baseline, the performance of our method improved by 0.349 in precision and 0.104 in the F1 score.