基于卷积神经网络的非合作目标两阶段位姿估计方法

A Two-Stage Pose Estimation Method for Noncooperative Targets Based on Convolution Neural Network

针对空间非合作目标在轨服务中的位姿估计问题,提出了一种基于卷积神经网络的两阶段相对位姿视觉估计算法.该算法在阶段1中结合位置回归和检测模块,将检测后的图像输入阶段2中,并对任务过程中的绕飞和接近两种情况分别设计姿态估计模型,绕飞时采用分类代替回归的间接方法,接近时采用直接回归方法估计姿态,实现了对非合作目标在轨服务过程的位姿估计.充足的消融实验验证了各阶段模型的有效性,仿真实验位置精度可达0.1836 m,姿态精度可达2.9489°,表明了基于卷积神经网络的单目视觉方法应用于非合作目标在轨服务中位姿估计的可行性.

A two-stage relative pose estimation algorithm based on convolutional neural network was proposed to solve the problem of pose estimation for space noncooperative targets in orbit service.The detection module was combined with translation regression module in the first stage,and the detected image was input into stage two.An attitude estimation model was designed for flight around and flight approach during the mission.The indirect method of classification instead of regression was used in flying around,and the direct regression method was adopted to estimate the attitude when approaching,so as to realize pose estimation of noncooperative targets in orbit service process.A large-scale dataset is introduced,which can be utilized as a benchmark for pose estimation methods.Abundant ablation studies verified the effectiveness of each module.The position accuracy could reach 0.1836 meters and attitude accuracy could reach 2.9489 degrees,which shows the feasibility of monocular vision method based on convolutional neural network to estimate the pose of noncooperative targets in orbit service.