Simple Single-Person Fall Detection Model Using 3D Pose Estimation Mechanisms

Abstract

The falling- and sliding-down (fall) accidents among the elderly are a major concern due to the potential to cause significant functional damage. This demands immediate medical care to prevent the injuries from progressing when the accident first happens. Although various technologies with wearables and vision systems that utilize artificial intelligence (AI) have been developed to detect falls, many AI models are complex and resource-intensive. Therefore, we propose the simple fall detection model (SFDM) that is computationally efficient with moderate accuracy. SFDM utilizes pose estimation to generate 3D human landmarks from live video streams. These landmarks are then processed to determine the cumulative and final status of fall or not fall by accumulating single detection results and applying a Run-Length Encoding (RLE) algorithm, with no training of AI. Our experiments are conducted with four case studies representing different fall or normal scenarios. Our results show that SFDM outperforms selected other models in per-frame floating-point operations, achieving the lowest model complexity. Future work will focus on refining the algorithm to improve accuracy, expanding the detection capability from single- to multi-person detection, and broadening the detection range with other daily living activities in elderly environments.

Keywords

• Artificial intelligence
• complexity
• fall detection
• floating-point operations
• pose estimation