摘要
针对头部姿态估计方法受特征提取限制导致姿态估计效果不佳的问题,提出使用深度卷积网络自动学习有效特征并进行分类的头部姿态估计方法。首先,利用DCNN非线性映射和自动提取图像结构信息的能力,设计一个深度卷积网络实现对姿态鲁棒特征的提取;然后,将提取的特征用于分类器训练并最终实现头部姿态估计。在Pointing’04和Face Pix数据库上的测试结果表明,本文设计的深度卷积网络能有效地进行特征学习,避免了人工设计特征的不足,与现有的基于人工设计特征方法相比,本文方法在2个数据库上达到的预测平均绝对误差分别为4.05°和2.04°,充分证实了本文算法的稳定性和可靠性。
In order to improve the estimation accuracy,a new method based on deep convolutional neural networks is proposed to automatically learn effective features and conduct classification for head pose estimation. First,a DCNN was designed and fulfilled,making use of DCNN's capability of nonlinear mapping and feature representation learning. Then,a classifier was trained based on the learned feature representations for the purpose of head pose estimation. Experimental results on the Pointing'04 and Face Pix databases showed that the designed DCNN could learn features automatically and efficiently for effective head pose estimation. The mean absolute errors on the two databases were 4. 05° and 2. 24°,respectively. The proposed DCNN based method outperforms existing hand-craft feature based methods in terms of both estimation accuracy and robustness.
出处
《四川大学学报（工程科学版）》
EI
CAS
CSCD
北大核心
2016年第S1期163-169,共7页
Journal of Sichuan University (Engineering Science Edition)
基金
国家重大科学仪器设备开发专项资助项目(2013YQ490879)
国家自然科学基金资助项目(61202160)
关键词
深度卷积网络
姿态估计
姿态分类
deep convolutional neural networks
head pose estimation
classification
