基于改进ResNeSt的头部姿态估计方法  

Head pose estimation method based on improved ResNeSt

作　　者：刘东 潘炼[1] LIU Dong;PAN Lian(College of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China)

机构地区：[1]武汉科技大学信息科学与工程学院,武汉430081

出　　处：《信息技术》2023年第5期101-106,114,共7页Information Technology

摘　　要：针对单目RGB图像进行头部姿态估计精度不高的问题,提出了一种具有尺度平移不变性的改进ResNeSt方法。该方法以卷积神经网络ResNeSt作为主干网络;为增强卷积网络对尺度平移的鲁棒性,参考传统图像处理方法,集成模糊滤波器于卷积网络降采样过程;最后引入多任务学习,将网络提取的特征图输出到混合角度分类分支,提升卷积网络分类精度。在公开数据集验证,结果表明,提出方法的头部姿态估计误差降低了18%,且在CPU上检测速度为0.075秒/张,具备较好的实时检测性能。In order to solve the problem that the method of using monocular RGB images for head pose estimation has low accuracy,an improved ResNeSt method with scale translation invariance is proposed.The method uses the Convolutional Neural Network(CNN)ResNeSt as the backbone network.In order to enhance the robustness of CNN to scale translation,a blur filter is integrated into the down-sampling process by referring to the traditional image processing method.Finally,multi-task learning is introduced,the feature map extracted from the network is output to the hybrid classification branches to improve the classification accuracy.In the public data verification,the results show that the proposed method can reduce the head pose estimation error by 18%,and the detection speed on CPU is 0.075 second per image,which has better real-time detection performance.

关 键 词：头部姿态估计 尺度平移不变性 ResNeSt 模糊滤波器 混合分类分支 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]