一种基于深度学习的头部姿态估计模型

Head Pose Estimation Model Based on Deep Learning

针对传统头部姿态估计网络存在空间结构信息易丢失问题,论文提出一种将胶囊网络与传统卷积神经网络相结合的头部姿态估计网络模型。该模型采用具有多级输出结构的传统卷积神经网络,将不同层级的空间结构信息和语义信息进行提取,同时利用胶囊网络能够充分保留特征信息的优点,将提取的特征进行编码,从而使其以胶囊的形式进行传递和输出,有效避免了空间结构信息丢失的问题。实验结果表明,论文提出的模型在AFLW2000和BIWI数据集上的平均绝对误差分别为5.68和4.33,进一步提高了对头部姿态估计的准确度,并在室内条件下对光照变化、遮挡等具有较好的鲁棒性。

Aiming at the problem that the spatial structure information of traditional head pose estimation network is easy to lose,this paper proposes a head pose estimation network model combining capsule network and traditional convolutional neural network. The model uses a traditional convolutional neural network with a multi-level output structure to extract spatial structure information and semantic information at different levels. At the same time,the advantages of the capsule network can fully retain the feature information and encode the extracted features,so that the features can transfer and output in the form of a capsule,effectively avoids the problem of loss of spatial structure information. The experimental results show that the average absolute errors of the proposed model on the AFLW2000 and BIWI datasets are 5.68 and 4.33,respectively,which further improves the accuracy of head posture estimation,and has good robustness on lighting changes and occlusion under indoor conditions.