基于YOLOv5和RepVGG的实时无标记点头部姿势估计方法  

Real-time Marker-free Head Pose Estimation Method Based on YOLOv5 and RepVGG

作　　者：蓝海倩 胡晓宇 张加宏[1] LAN Hai-qian;HU Xiao-yu;ZHANG Jia-hong(College of Electronic and Information Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;Zhongke Nanjing Intelligent Technology Research Institute,Nanjing 211135,China)

机构地区：[1]南京信息工程大学电子与信息工程学院,江苏南京210044 [2]中科南京智能技术研究院,江苏南京211135

出　　处：《中国电子科学研究院学报》2025年第1期41-47,共7页Journal of China Academy of Electronics and Information Technology

基　　金：国家重点研发计划(2022YFB3205903)。

摘　　要：为了在实际应用中解决遮挡、模糊和极端姿态等情况会显著影响标记点检测的性能,进而影响基于标记点检测的头部姿势估计的准确性和鲁棒性的问题,文中提出一种基于YOLOv5和RepVGG骨干网络的实时无标记点头部姿势估计方法。该方法主要是利用YOLOv5的目标检测能力和RepVGG-B1g4骨干网络的高效卷积特性,并在特征融合模块引入非局部自注意力机制提升模型特征表示能力,再结合连续的6D旋转矩阵表示以避免万向锁问题,并在模型训练过程中提出一种多重回归损失函数,实现高效且稳健的头部姿势角度估计。实验结果表明,文中所提方法在AFLW 2000测试集上头部姿势角度估计取得了4.32°的平均绝对误差。In order to solve the problem that situations such as occlusion,blurring and extreme poses can significantly affect the performance of marker point detection in practical applications,which in turn affects the accuracy and robustness of head pose estimation based on marker point detection,a real-time marker-free head pose estimation method based on YOLOv5 and RepVGG backbone network is proposed.This approach utilizes the object detection capabilities of YOLOv5 and the efficient convolutional properties of the RepVGG-B1g4 backbone network.It incorporates a non-local self-attention mechanism in the feature fusion module to enhance the model’s feature representation ability,combined with a continuous 6D rotation matrix representation to avoid gimbal lock issues,and a multi-regression loss function is introduced during model training to achieve efficient and robust head pose angle estimation.Experimental results show that this method achieves an average absolute error of 4.32°in head pose angle estimation on the AFLW2000 test set.

关 键 词：头部姿势估计 无标记点 YOLOv5 RepVGG 6D旋转表示 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]