Deep Multi-Task Locomotion-Invariant Gesture Recognition with FMCW Radar

Authors

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

Abstract

Recognizing hand gestures with FMCW mmWave radar becomes substantially harder when users are walking, because locomotion-induced Doppler and clutter interfere with gesture signatures. This paper proposes a multi-task deep framework for locomotion-invariant dynamic gesture recognition. An adaptive interference suppression (AIS) pipeline maps Range–Doppler cubes to Range–Angle Projection (RAP) tensors, which are processed by a modified MambaVision backbone. A multi-task head jointly predicts gesture class and locomotion state, with supervised contrastive regularization on gesture-only positives and a two-phase training schedule that first emphasizes the auxiliary locomotion objective and then fine-tunes with a reduced auxiliary weight. Experiments on the DSMU-RAD / LAG-ID radar dataset demonstrate robust gesture recognition across stationary and walking conditions.