A lightweight AI system for real-time elderly fall detection

Elderly fall detection is a critical healthcare challenge that demands accurate, reliable, and real-time solutions suitable for deployment in sensitive environments, such as nursing homes. Although significant advances have been made in sensing technologies and machine learning, the role of data preprocessing, particularly feature normalization, remains underexplored, despite its substantial impact on model performance. This paper presents a lightweight vision-based fall detection system that has been deployed and evaluated in a real-world elderly care facility, addressing a key limitation of previous work that often lacks real-life validation. Central to the approach is a systematic investigation into the effect of eight normalization techniques on the performance of four representative classification models. The results show that appropriate normalization, particularly Min-Max and Z-score normalization, leads to substantial improvements in classification accuracy and F1 score, while also reducing training time by a factor of 45. Based on these findings, the proposed system incorporates min-max normalization into a hybrid architecture that combines SSD MobileNet V2 with geometric heuristics for real-time fall detection. The system is privacy-aware, non-intrusive, and practical for real-world healthcare deployment.