AttFallNet : Attention-Guided Fall Detection System Using Improved YOLOv8 Network
Li Ye, Jing Lei · 2024
Fall detection is essential for ensuring the safety of elderly individuals, particularly in environments where immediate assistance may not be available. Despite advancements in Convolutional Neural Networks (CNNs) for fall detection, existing models often face challenges in the accuracy of detection falls due to the complexity and variability of such events. In response, we propose an improved version of the YOLOv8 model called AttFallNet to improve fall detection accuracy and reliability. The first enhancement involves replacing the original Spatial Pyramid Pooling Fast (SPPF) block with an Atrous Spatial Pyramid Pooling (ASPP) block, which captures multi-scale features more effectively and improves the model’s ability to detect diverse fall patterns. The second enhancement introduces attention blocks, which refine the network’s focus on relevant spatial information, thereby enhancing detection performance. We train and validate AttFallNet on a publicly available CAUCAFall dataset, and our experimental results demonstrate that our proposed AttFallNet has increased by 0.62% to 99.21% in accuracy compared to YOLOv8. Also, AttFallNet achieves a detection precision of 97.9%, recall of 99.1%, and mean Average Precision (mAP) of 99.5%, thus outperforming existing baseline models and providing a robust and reliable solution for fall detection. This work underscores the potential of AttFallNet to reinforce safety systems in elderly care, offering a new benchmark for fall detection technology.