文献类型：期刊文献

英文题名：Dual-Branch Gaze Estimation Algorithm with Gaussian Mixture Distribution Heatmaps and Dynamic Adaptive Loss Function

作者：Dai, Songyin[1]; Zhang, Chaoran[1,3]; Xu, Cheng[1,3]; Yan, Chao[1]; Huang, Jie[2]

第一作者：代松银

机构：[1] Beijing Key Laboratory of Information Service Engineering, Beijing Union University, Beijing, 100101, China; [2] Beijing Information Technology College, Institute of Artificial Intelligence, Beijing, 100015, China; [3] Beijing Qiangqiang Yuanqi Technology Co., Ltd, Beijing, 101125, China

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2025

卷号：34

期号：5

起止页码：433-446

外文期刊名：Journal of Beijing Institute of Technology (English Edition)

收录：EI(收录号：20260219900415);Scopus(收录号：2-s2.0-105027204546)

语种：英文

外文关键词：Angular distribution - Convolutional neural networks - Dynamics

摘要：Gaze estimation, a crucial non-verbal communication cue, has achieved remarkable progress through convolutional neural networks. However, accurate gaze prediction in unconstrained environments, particularly in extreme head poses, partial occlusions, and abnormal lighting, remains challenging. Existing models often struggle to effectively focus on discriminative ocular features, leading to suboptimal performance. To address these limitations, this paper proposes dual-branch gaze estimation with Gaussian mixture distribution heatmaps and dynamic adaptive loss function (DMGDL), a novel dual-branch gaze estimation algorithm. By introducing Gaussian mixture distribution heatmaps centered on pupil positions as spatial attention guides, the model is enabled to prioritize ocular regions. Additionally, a dual-branch network architecture is designed to separately extract features for yaw and pitch angles, enhancing flexibility and mitigating cross-angle interference. A dynamic adaptive loss function is further formulated to address discontinuities in angle estimation, improving robustness and convergence stability. Experimental evaluations on three benchmark datasets demonstrate that DMGDL outperforms state-of-the-art methods, achieving a mean angular error of 3.98° on the Max-Planck institute for informatics face gaze (MPI-IFaceGaze) dataset, 10.21° on the physically unconstrained gaze estimation in the wild (Gaze360) dataset and 6.14° on the real-time eye gaze estimation in natural environments (RT-Gene) dataset, exhibiting superior generalization and robustness. ? (2026). All right reserved.