Fall Detection by Pose Estimation with Thermal Infrared Images for Elderly in Daily Life

Falls present a significant risk to elderly individuals in their daily life, particularly during nighttime activities such as getting in and out of bed or using the bathroom. Accurate and rapid detection of falls can enable caregivers to intervene promptly, thereby reducing potential harm. To address issues of invasiveness, reliability, and privacy associated with existing methods, an automatic fall detection method based on thermal infrared images is proposed. The approach utilizes a human pose estimation network (AlphaPose) to extract human keypoints, which are then fed into a spatio-temporal graph convolutional network (ST-GCN) to detect falls. Initially, to enhance the performance of AlphaPose, a YOLOv7-tiny framework is integrated, augmented with a squeeze-and-excitation self-attention mechanism, and traditional upsampling is replaced with content-aware reassembly of features (CARAFE). Additionally, a low-complexity Kalman filter is employed for target tracking, ensuring stable fall detection. Experimental results demonstrate that, for images with resolutions ranging from 416×416 to 96×96, the improved model achieves a high mean average precision (mAP) in 2D pose estimation. The proposed method ensures immediate response during fall events, achieving 100% accuracy while maintaining a low computational complexity.