Fine-Grained Fall Type Recognition Using Low-Cost Wearable IMU via Robust Motion Perception and Elastic Sequence Classification

For fall detection in elderly and disabled populations, existing studies have primarily focused on identifying the occurrence of falls, while lacking fine-grained classification of fall patterns— such as forward/backward falls, slips, trips, and rotational falls, which is crucial for injury assessment and improving the efficiency of medical response. Such movements often exhibit short-term abrupt changes, nonlinearity, and partially constrained dynamics, making it challenging to achieve real-time robust recognition using low-cost wearable IMUs. To meet these demands, this paper proposes a fall recognition framework based on a low-cost wearable watch-type IMU, integrating dynamic adaptive error compensation, strong state tracking, and nonparametric temporal alignment-based classificationachieving high-precision, low-latency recognition of fall patterns. Specifically, an improved adaptive joint update algorithm (ZUPT + ZARU) based on zero-velocity and zero-angular-rate is first designed using a sliding-variance window adaptively adjusted according to motion stability, enabling real-time drift suppression across diverse fall types. Second, an improved Adaptive Strong Tracking Kalman Filter (ASTKF) is proposed to enhance responsiveness to sudden state transitions through a residual-driven covariance adjustment mechanism. Finally, an elastic alignment capability of Dynamic Time Warping (DTW) is integrated with the nonparametric adaptability of the KNN classifier to handle classification challenges caused by transient motion dynamics. Experimental results demonstrate that the proposed method effectively mitigates localization drift in low-cost devices, improves motion data accuracy, and enables efficient recognition of fine-grained motion patterns.