Composite Multi-Stage Transfer Learning with Cross-Scene Adaptation for Personalized In-Vehicle Gaze Estimation

Transferring gaze estimation models from mobile terminals to in-vehicle driving scenarios faces challenges from differences in device form, viewpoint, and imaging conditions. To address this, this study proposes a cross-scene adaptation framework that integrates composite multi-stage transfer learning with distributed multi-camera fusion for personalized in-vehicle gaze estimation. The framework builds upon a multi-level information element fusion baseline (MIEF) and explores three transfer paths (CMP1-CMP3) for progressive personalized adaptation from a large-scale mobile source domain to an in-vehicle target scenario. A distributed multi-camera platform is established, producing gaze calibration data (GCD) for cross-session training and driving gaze and road scene data (DRD) for visualization validation. A Multi-Camera Fusion Model (MCF) is designed with two paradigms, MCF-Lite and MCF-Plus, to balance accuracy improvement and computational complexity. Experimental results demonstrate that the composite multi-stage transfer learning path CMP3 achieves an average error of 0.277 across camera views, with a best single-camera mean error of 0.244. Building on this, the distributed MCF-Plus model with the full C1 + C2 + C3 configuration further reduces the mean error to 0.160 and the standard deviation to 0.127, with 96.88% of predicted gaze points falling within the target calibration block (Error < 0.5). The proposed method provides a feasible technical pathway for personalized adaptation of gaze estimation models from mobile-terminal source domains to in-vehicle multi-camera target scenarios, and offers preliminary experimental evidence for gaze behavior perception and driving gaze heatmap generation in real environments.