6-DoF Pose Estimation for Axisymmetric Objects Using Deep Learning with Uncertainty

Abstract

Before space debris can be removed efficiently, its 6-DoF poses (positions and attitudes) need to be estimated accurately from observed images with high resolution. Further, if the debris is axisymmetric, such as the remains of a multistage rocket, or if part of the debris cannot be seen due to optical conditions, it is considerably more difficult to estimate its parameters. If some parameters cannot be estimated for some reason, all parameters may be affected because each parameter in Euler angle and quaternion has an interdependency and the solution will not be determined uniquely. This research proposes methods that obtain the solution by decomposing the quaternion into the direction and rotation based on the forward direction so that direction and rotation parameters can be estimated independently. Moreover, this research was able to adaptively improve accuracy based on a threshold of uncertainty by adding an uncertainty value to each parameter. When the estimated parameters likely having error values that exceed 2% based on uncertainty value are deleted, estimated error of parameter $x, y, z$ (position), $n_{x}, n_{y},n_{z}$ and $\theta_{z}$ (attitude) were 1.25%, 1.35%, 3.76%, 2.27%, 2.64%, 3.06%, and 18.32% respectively.