Head pose estimation (HPE) under infrared imaging has become more and more common in the human-computer interaction. In this paper, we proposed a novel HPE with convolutional neural network and established a precise head pose database under 5° angle (HPD5A) for human attention recognition. Specially, the HPD5A database includes 729 infrared head pose images from different subjects with and without glasses, which corresponds to drivers who wear glasses or do not wear glasses. To verify the availability and usability of the HPD5A database, the benchmark evaluations are performed on our database using traditional standard HPE classification methods with and without principal component analysis. The methods include linear discriminant analysis, K-nearest neighbor, random forest and Naïve Bayes classifiers. We also design and implement a convolutional neural network architecture as one of elementary assessments. All the results are provided for future reference. The developed deep learning technique could obtain the state-of-the-art performance on the HPD5A database. This database will certainly help in the development of model for infrared HPE and be beneficial to the attention recognition in human-computer interaction system.

红外成像下的头部姿态估计（HPE）在人机交互中已越来越普遍。在本文中，我们提出了一种具有卷积神经网络的新型HPE，并建立了一个5°角下的精确头部姿势数据库（HPD5A），用于人类注意力识别。特别是，HPD5A数据库包含来自有眼镜和无眼镜的不同对象的729个红外头部姿势图像，对应于戴眼镜或不戴眼镜的驾驶员。为了验证HPD5A数据库的可用性和可用性，使用传统的标准HPE分类方法在有和没有主成分分析的情况下对我们的数据库进行基准评估。该方法包括线性判别分析，K-最接近的邻居，随机森林和朴素贝叶斯分类器。我们还将设计和实现卷积神经网络体系结构作为基本评估之一。所有结果均提供以供将来参考。所开发的深度学习技术可以在HPD5A数据库上获得最先进的性能。该数据库必将有助于红外HPE模型的开发，并有利于人机交互系统中的注意力识别。