Deep Gesture Recognition Under Data Loss

Abstract

Hand gesture recognition (HGR) using radar sensors enables intuitive, contact-free human–computer interaction. However, real-world deployment is hindered by data loss (DL) induced by interference from nearby radars and environmental obstructions. To address this challenge, we propose a deep learning-based sequence modeling framework that is robust to missing or corrupted data. The system employs an encoder–decoder network to reconstruct clean latent representations from impaired inputs, followed by a scale-shift module that aligns features with gesture categories for accurate classification. Experimental results demonstrate strong resilience, maintaining stable performance with up to 35% DL. Under 20% DL, the system achieves 98.50% recognition accuracy. To evaluate generalization, we tested the model on a separate dataset collected in a new environment with an unseen user and real interference from one or two additional mmWave radars. These corrupted samples were used exclusively for testing, while DL was simulated during training. The proposed system maintained 96.61% accuracy under these adverse conditions, confirming its robustness and reliability in practical, data-degraded scenarios.