Advanced Centroid State Estimation and Instability Detection Framework Leveraging UKF and AdvSAE-LSTM

With the growing aging population, fall detection has become a significant research focus in recent years. However, most existing studies have not effectively improved the accuracy of centroid state estimation in human instability, particularly in terms of multi-source information fusion. To address this, we propose a centroid state estimation method based on the Unscented Kalman Filter (UKF), integrating data from the five-link model and trunk IMU sensors, significantly enhancing both the accuracy and robustness of centroid state estimation. Experimental results demonstrate that the proposed method achieves estimation errors of 0.0099 m and 0.0226 m/s for centroid displacement and velocity in the x-direction, and 0.0006 m and 0.0099 m/s in the y-direction, indicating high accuracy. To overcome the reliance of traditional supervised learning methods on large amounts of labeled data for instability state estimation, we propose an unsupervised learning-based instability state estimation method, featuring the AdvSAE-LSTM model. Data from different sensors are preprocessed and independent training datasets are constructed, after which the feature maps from each sensor are concatenated to form a global feature map. Compared to baseline methods, the proposed approach significantly improves instability state detection performance. Specifically, when the SAS threshold is set to 0.02, both the False Positive Rate (FPR) and False Negative Rate (FNR) are 0%, with a detection delay of only 77.16 ms, achieving highly efficient and accurate estimation of human instability states.