Non-Visual Human Eye Gaze Tracking Using Radar and Artificial Intelligence for Robust Driver Monitoring Systems

Abstract

Estimating gaze direction from in-cabin sensors is central to driver distraction monitoring. Existing methods rely on cameras, which degrade under adverse lighting and collect identifiable facial imagery. Millimetre-wave radar avoids both limitations, and prior work at 60 GHz has demonstrated sensitivity to eye-region micro-motion such as blinks. Whether radar can support directional gaze estimation, rather than binary event detection, remains an open question. This thesis develops a radar-only gaze estimation pipeline built on a 60 GHz FMCW sensor. Amplitude and inter-receiver phase-difference cues are extracted from the radar return, and candidate eye-movement events are detected from the radar signal alone, removing the need for camera or stimulus timing at inference. A lightweight dual-stream temporal architecture, DualStreamGazeNet, encodes the two cue types separately, fuses them through cross-modal self-attention, and produces both a direction label and a continuous gaze angle through a modality-aware hierarchical classifier and an auxiliary regression head. Experiments on a multi-session dataset show that the model achieves 85.7% balanced accuracy for four-direction classification with azimuth and elevation errors of 7.82◦ and 3.15◦. Under strict cross-session evaluation, few-shot calibration with three to five labelled events per direction yields 82.1% balanced accuracy, demonstrating that the learned representation generalises effectively with minimal target-session adaptation. These results establish that close-range mmWave radar is a viable modality for event-level gaze estimation and can serve as a privacy-preserving, illuminationindependent complement to driver monitoring system.