遮挡与几何感知模型下的头部姿态估计方法

Head Pose Estimation Method Under Occlusion and Geometric Perception Models

头部姿态估计在人机交互、辅助驾驶等应用中起重要作用,但因受到光照变化、部分遮挡、不同面部外观差异等因素的影响,导致头部姿态估计的准确率不高。提出一种基于遮挡和几何感知模型的头部姿态估计方法,在预处理阶段采用多任务卷积神经网络进行人脸检测,减小背景环境的干扰,并进行图像增强操作以减小光照变化带来的影响。设计面部遮挡感知网络感知人脸的遮挡区域,从而提取信息量丰富的未遮挡面部特征。为充分利用面部的几何信息,采用堆叠胶囊自编码器对人脸各部分的姿态和位置进行编码,得到面部各部分间的几何关系。实验结果表明,该方法在AFLW2000数据集和BIWI数据集上的平均绝对误差分别为3.91和3.55,能有效提高头部姿态估计准确率,在复杂环境下的鲁棒性较好。

Head posture estimation plays a crucial role in human-computer interaction and assisted driving.However,lighting variations,partial occlusion,and differences in face appearances degrade the accuracy of head pose estimation.Therefore,in this study,a head pose estimation method based on occlusion and geometric perception models is proposed.During the pre-processing phase,a Multi-Task Convolutional Neural Networks(MTCNN)is used for face detection to reduce the interference of the background environment,and then,an image enhancement operation is performed on the captured image to reduce the effect of illumination changes.A facial occlusion perception network is designed to perceive the occluded area of the face to extract unobstructed facial features with rich information.The posture and position of each part of the face are encoded using the Stacked Capsule Autoencoder(SCAE)to completely utilize the geometric information of the face,so that geometric relationship of each part of the face is obtained.The experimental results based on the AFLW2000 and BIWI datasets indicate that the Mean Absolute Error(MAE)values of the proposed method are 3.91 and 3.55,respectively,which can effectively help improve the accuracy of head pose estimation and its robustness in complex environments.