End-to-end infrared fall detection with privacy preservation
Abstract
To address the issues of privacy leakage and nighttime ineffectiveness associated with visible light monitoring in home-based elderly care scenarios, this study employs infrared thermal imaging technology to achieve privacy-preserving human fall detection. Due to the lack of publicly available high-quality indoor infrared fall datasets covering diverse fall postures and confounding daily actions, an indoor data acquisition environment was established to construct a dataset encompassing multiple perspectives and complex actions. We adopt an end-to-end Real-Time Detection Transformer architecture, and proposes a tailored infrared fall detection scheme for home-based elderly care scenarios. By leveraging its global self-attention mechanism and Non-Maximum Suppression-free inference characteristics, the model effectively adapts to the blurred and low-contrast features of infrared images, thereby realizing all-weather, privacy-preserving fall detection. Experimental results demonstrate that the proposed system achieves an mAP@0.5 of 0.975 and an optimal F1 score of 0.980 on the self-constructed dataset. It can effectively distinguish genuine falls from easily confused daily actions, satisfying the requirements for 24-hour real-time monitoring in highly private areas such as bedrooms and bathrooms.