Abstract
Falls are a leading cause of injury among older adults, with research indicating that they often fall due to certain individual biomechanical factors. Therefore, real-time individual fall risk assessment is essential for designing more effective fall prevention programs and developing advanced home-based training solutions. However, existing methods for fall risk assessment either raise privacy concerns due to sensors installed in the environment or require multiple wearable devices, limiting their practicality for home-based applications and long-term monitoring. In this paper, we introduce EgoFall, a real-time privacy-preserving fall risk assessment system. EgoFall utilizes a chest-mounted commercial tracking camera and a carefully designed data pre-processing pipeline to acquire the ego-body motion data of the subject. The data is then fed to a lightweight CNN-Transformer model for fall risk assessment. To evaluate the proposed method, we establish the EgoWalk dataset, which includes four walking patterns: normal, anterior-posterior instability, medial-lateral instability, and combined instability. Experimental results show that EgoFall achieves an accuracy exceeding 95% on the EgoWalk dataset, outperforming baseline methods while maintaining low computational complexity. Additionally, a series of ablation studies explore the impact of fine-tuning data and error analysis, further highlighting EgoFall's practicality in real-world applications.