6-DoF Pose Estimation from Stereo LiDAR of Actual Machine using Deep Learning

Abstract

We adopted two LiDARs as input instead of an optical camera in order to achieve highly-accurate pose estimation at short distances for our space debris removal mission. To overcome the inability to obtain such point clouds of debris on orbit, we developed a simulator and a hardware emulator. There has been no research on suppressing the domain shift of simulated data and emulated data and estimating pose with high accuracy and robustness from stereo point clouds. We maximized the representation of voxels and minimized the size by normalizing the point clouds' extent rather than the space. This provided fixed-length data for a small input suitable for a neural network. In addition, we gave the neural network eight vertices indicating the position of the point clouds. This interpolates the positions of point clouds that have become dimensionless by normalization. We performed environmental randomization to suppress domain shifts. Voxels were randomly removed when input to a neural network based on dropout rate. Making the dropout rate variable rather than fixed improved the accuracy of attitude estimation. This paper also proposed a method to effectively process stereo point clouds by neural networks. The estimation accuracy was improved by a method that extract features individually from stereo point clouds upstream and interfere its features downstream to infer the pose. The dropout layer was often used for approximate Bayesian inference and environment randomization. We performed pose estimation on simulation and emulation point clouds with the same pose labels. With 24 patterns of data, the simulation results had estimated errors of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$x, y, z, n_{x}, n_{y}, n_{z}$</tex> , and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\theta$</tex> , which were 0.59%, 0.36%, 0.97%, 1.22%, 1.05%, 0.02%, and 3.47%, respectively. For the emulation results, the estimated errors of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$x, y, z, n_{x}, n_{y}, n_{z}$</tex> , and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\theta$</tex> were 0.91%, 0.58%, 0.93%, 4.78%, 2.63%, 0.05%, and 3.44%. There were no large errors in the results from simulated or emulated point clouds, confirming that domain shift suppression was successful. We experimented using the uncertainty output by Bayesian inference to eliminate pose estimation results from noisy inputs. We removed the estimation results, which had high uncertainty values. The number accepted was 9 out of 24. As a result, the estimated errors of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$x, y, z, n_{x}, n_{y}, n_{z}$</tex> , and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\theta$</tex> were 0.52%, 0.52%, 1.11%, 1.10%, 2.02%, 0.03%, and 3.39%, respectively. Although the amount of data was reduced by about half, the accuracy of all parameters was improved.