【摘要】 患者头部姿态估计技术是神经外科手术机器人自主化与智能化感知的关键技术之一。本研究利用基于数据驱动的深度学习方法帮助神经外科手术机器人估算患者头部姿态，为神经外科手术的智能化打下基础。首先，建立患者头部姿态估计任务的基本数学关系；其次，给出一种高效、鲁棒的头部姿态标注方法，解决面部特征缺失情况下的2D头部图像姿态标注问题；之后，通过采集机器人视角下神经外科手术场景照片，构建共包含79个手术场景、合计4 301张照片的患者头部姿态估计数据集；最后，研究HopeNet深度神经网络在患者头部姿态估计问题上的适用性，并通过分布式标签、头部区域自适应裁剪、旋转数据增强以及新提出的旋转速率损失函数等方法提高模型性能。进行网络训练和评价，在包含10个手术场景、386张照片的同源测试集1上，基于单视角的姿态估计在偏航角、滚动角、俯仰角等3个方向平均可以达到±12.76°的预测误差；在8个手术场景、229张照片的异源测试集2上，在上述3个方向平均可以达到±13.41°的预测误差。结果表明，提出的模型能准确估计患者头部姿态，且提出的优化方法可以有效提升算法精度，并提高模型的泛化能力。更多还原

【Abstract】 Patient head pose estimation is one of the key technologies for autonomous and intelligent perception of neurosurgery robots. This paper aimed to use data-driven deep learning method to help the neurosurgery robot to estimate the patient′s head posture, laying the foundation for the intelligence of neurosurgery. This paper firstly established the basic mathematical relationship of the patient head pose estimation task. Next, an efficient and robust head pose labeling method was proposed to solve the problem of 2 D head image pose labeling in the absence of facial features. After that, by collecting the neurosurgery scene photos from the perspective of the robot, a patient head pose estimation dataset containing a total of 79 surgical scenes and a total of 4 301 photos was constructed. Finally, the applicability of the HopeNet deep neural network in the patient head pose estimation problem was studied, and methods including cropping, rotation data augmentation, and our newly proposed rotation rate loss function improved the model performance. For the network training and evaluation, on the homologous test set 1 containing 10 surgical scenes and 386 pictures, the pose estimation based on a single perspective could reach an average of ±12.76°in three directions including yaw angle, roll angle, and pitch angle; on the heterogeneous test set 2 of 8 surgical scenes and 229 photos, the average prediction error of ±13.41° could be achieved in the three directions. The results showed that the proposed model could accurately estimate the patient′s head pose, and the proposed optimization methods could effectively improve the accuracy of the algorithm and improve the generalization performance of the model.更多还原