Temporal Shift Module (TSM) Based Automatic Fall Detection with Bounding Box Grounding

Abstract

Fall detection is one of the important tenets of remote geriatric care operations. Fall is one of the main causes of injury in old individuals leading to fractures, concussions, and different issues that might lead to prompt demise. In a world increasingly making the elderly live in isolation, accurate and real-time detection of falls is very important to remote caregivers to be able to provide timely medical assistance. Recent advancements in vision-based technologies have got promising results; however, these models are often trained on acted datasets and their appropriateness for application in the wild is not well established. In this paper, we propose a vision-based fall detection mechanism that improves the accuracy of in-the-wild complex events. The proposed system is built leveraging Temporal Shift Module (TSM) with a bounding box grounding (BBG) approach for accurate Region Of Interest (ROI) sequence generation when sudden deformation in the shape is observed. Compared to the general 3D CNN based approaches, the proposed model achieves better accuracy while maintaining the level of computational complexity at that of the 2D CNN models. The proposed approach demonstrates promising performance on both acted and in-the-wild datasets.