TKDCoach: A Jerk-Aware IoT System for Human Activity and Skill Level Recognition via Large Language Models

Abstract-The rapid advancement of the Internet of Things (IoT) has enabled data-driven sports analytics; however, real-time Taekwondo assessment remains limited by complex motion signals, privacy concerns, and restricted edge-device capacity. Existing wearable systems relying on handcrafted features or shallow models struggle to capture skill difference among athletes. This paper presents TKDCoach, an IoT-based and interpretable AI-based coaching framework integrating jerk-aware sensing with Large Language Model (LLM)-driven semantic reasoning. Multiple Inertial Measurement Unit (IMU) nodes placed on the waist, thigh, and shank record six-axis motion for five representative Taekwondo techniques. A jerk-enhanced feature extractor encodes fine-grained variations into a compact codebook, while a wavelet-guided loss preserves both smooth trajectories and transient impacts. The resulting motion tokens are processed by a cloud LLM to perform semantic evaluation and generate feedback, while experiments verify the method on public Human Activities Recognition (HAR) benchmarks and a self-collected Taekwondo dataset. TKDCoach delivers scalable, interpretable, and privacy-preserving skill analysis, effectively addressing the above challenges.