Learning to Track: Deep Association for Multipath-Resilient In-Air Writing with D-band FMCW Radar

This paper presents a framework for multipath-robust in-air writing using two co-located D-band FMCW radar sensors. While millimeter-wave radar offers privacy-preserving, high-resolution sensing, trajectory reconstruction is often compromised by multipath and the extended-object reflections. We address these challenges through a cascaded pipeline. First, we employ DBSCAN clustering to aggregate the sparse CFAR detections obtained from the range-Doppler maps, mitigating the jitter associated with naive peak detection. Second, we introduce a Deep Association logic, where a weakly supervised Bidirectional LSTM network learns to track the hand by distinguishing valid motion patterns from multipath reflections based on temporal history. Third, position and velocity states are estimated via an Extended Kalman Filter. Experimental results demonstrate that this architecture effectively suppresses multipath reflections, yielding smooth, legible ``digital ink'' suitable for human–computer interaction.