End-to-End Motion-Robust Gesture Recognition from Raw FMCW Data

Authors

Amin Kajbaf, Ehsan Yazdian, Mohammad Ali Akhaee, Ramin Toosi, Saeed Gazor

Abstract

Most radar gesture systems rely on handcrafted Range–Doppler or spectrogram features and are evaluated mainly on stationary users, limiting robustness when people gesture while walking. This paper studies end-to-end motion-robust gesture recognition that operates directly on raw complex intermediate-frequency (IF) FMCW radar measurements, without explicit FFT-based preprocessing. Using the RG-Motion dataset of 7,138 walking-gesture recordings from 60 participants across indoor and outdoor environments (12 dynamic gesture classes), we develop and evaluate deep models for subject-disjoint recognition under walking-induced motion interference, targeting robust contactless human–computer interaction with mmWave radar.