Micro-TsNN: A Lightweight End-to-End Gesture Recognition Framework for Millimeter-Wave Radar Based on Knowledge Distillation

Millimeter-wave radar-based dynamic gesture recognition holds significant potential for enabling non-contact human–computer interaction. However, existing recognition models often exhibit large computational footprints that hinder deployment on resource-constrained edge devices. To overcome this challenge, this paper proposes Micro‑TsNN (Micro Time-series Neural Networks), a lightweight end-to-end gesture recognition framework based on knowledge distillation. The framework directly processes intermediate‑frequency radar echo signals by reshaping them into a structured 3D tensor (Time × Sequence × Channel), thereby avoiding complex signal preprocessing pipelines. Architecturally, it integrates temporal feature extraction with inter‑frame self‑attention to capture motion dynamics and suppress environmental interference. We employ a response‑based knowledge distillation strategy, in which the full Self Attention Time‑series Neural Network (Atten‑TsNN) acts as the teacher model to guide the training of its compact, submodule‑based student counterpart, Micro‑TsNN. Experimental evaluation on a dynamic gesture dataset containing 20% randomized interference shows that Micro‑TsNN achieves 98.12% recognition accuracy with a model size of only 0.64 MB. Compared with directly trained lightweight baselines, it improves accuracy by 1.58% while reducing model parameters by 66.9%, confirming its suitability for real‑world embedded edge deployment.