39 / 2025-03-28 16:25:56

Driver Gaze Estimation Method Based on Improved Multi-Head Attention

Gaze Estimation,Human-Computer Interaction,Neural Network,Swin Transformer

To enhance driver gaze estimation by fully lever aging ocular features, this study proposes a parallel dual branch network that integrates an improved multi-head attention backbone with an eye feature extraction branch. To overcome challenges such as inadequate spatial information capture, in efficient multi-scale feature fusion, and underutilized channel information, a Gaze Optimization Module and Hierarchical Feature-Axis Convolution are introduced. The method follows a structured process: first, preprocessed face and eye images are input into the model. The enhanced multi-head attention backbone processes facial features, while a CNN-based branch extracts ocular features. Subsequently, an attention cross module and dual-feature aggregation module refine feature representa tion, ensuring effective fusion of global and local information. Finally, the driver’s gaze direction is estimated. Experiments on the MPIIGazeFace dataset show that the proposed method outperforms existing networks by reducing estimation errors, demonstrating its effectiveness in gaze prediction.