<p>Head pose estimation is a challenging and critical research topic, with existing models still facing significant challenges. First, common representations for head pose estimation exhibit discontinuities. Second, recognition rates are low in complex scenes, and models tend to have high parameter counts and substantial computational demands. To solve these problems, this paper proposes an unconstrained head pose estimation model based on bilinear attention. We introduce a 6D rotation matrix for attitude angle representation and a P-Ghost module to enhance the GhostNetV2 lightweight framework for feature extraction. A bilinear attention network is also introduced to integrate spatial and channel information, enabling the model to learn feature correlations, prioritize key channels, and suppress redundant ones. Multiple loss function strategies are also introduced to improve the model’s accuracy. The proposed network model undergoes extensive testing on three datasets, with experimental results showing superior performance in head pose estimation.</p>