Abstract
Falls in health and social care require privacy-preserving and cost-effective monitoring solutions to ensure safety. Low-resolution infrared (IR) array sensors provide a promising balance between invasiveness and information content; however, their performance can be strongly affected by ambient temperature variations and thermal interference. In this work, an environment-adaptive fall-detection system was developed to achieve cross-domain robustness and embedded real-time feasibility. A 24×32 MLX90640 IR array integrated with an ESP32 microcontroller captures thermal sequences at 8 fps. The processing pipeline includes temperature-based filtering, background removal based on a U-Net architecture for isolating human motion, Farneback optical flow extraction, and a lightweight CNN–BiLSTM–Attention network for spatiotemporal modeling. Experiments conducted across three thermally distinct indoor environments (11,148 samples, six activity classes) demonstrate that the U-Net-based background subtraction significantly enhances model robustness, achieving a six-class classification accuracy of 86.31% and reaching 98.27% in a simplified three-class (normal/falling/empty) configuration. Ablation and profiling analyses confirm the contribution of each processing stage while maintaining a compact model (≈0.83 M parameters, ≈3.17 MB) operating in real time at 8 fps on the ESP32–MLX90640 platform.
Funding Information
  • National Science and Technology Council, Taiwan (111-2622-E-A49-004)
  • National Science and Technology Council, Taiwan (114-2221-E-A49-172)