Pose Estimation of Uncooperative Spacecraft Using Monocular Vision and Deep Learning PDF Download

Author: Sumant Sharma
Publisher:
ISBN:
Category :
Languages : en
Pages :

Get Book Here

Book Description
This dissertation addresses the design and validation of new pose estimation algorithms for spaceborne vision-based navigation in close proximity to known uncooperative spacecraft. The onboard estimation of the pose, i.e., relative position and attitude, is a key enabling technology for future on-orbit servicing and debris removal missions. The use of vision-based sensors for pose estimation is particularly attractive due to their low volumetric and power requirements, particularly in comparison to sensors such as LiDAR. Past demonstrations of this technology have relied on various combinations of cooperative use of fiducial markers on the target spacecraft, frequent inputs from the ground control stations, and post-processing of the images on-ground. This research overcomes these limitations by developing novel pose estimation methods that take as input a single two-dimensional image and a simple three-dimensional (3D) wireframe model of the target. These methods estimate the pose without requiring any a-priori pose information or long initialization phases. The first pose estimation method relies on a novel feature detection algorithm based on the filtering of the weak image gradients to identify the edges of the target spacecraft in the image, even in the presence of the Earth in the background. As compared with state-of-the-art feature detection-based methods, this method is shown to be more accurate and computationally faster through experiments on flight imagery. The second pose estimation method leverages modern learning-based algorithms by using a convolutional neural network architecture to classify the pose of the target spacecraft in the image. As compared to the feature detection-based methods, this method is shown to be more robust and two orders of magnitude computationally faster during inference using experiments on synthetic imagery. The third pose estimation method, the Spacecraft Pose Network (SPN), combines the higher accuracy potential of feature detection-based methods with the higher robustness and computational efficiency of learning-based methods. The SPN method achieves this by integrating a convolutional neural network with a Gauss-Newton algorithm. The use of a convolutional neural network allows SPN to implicitly perform feature detection without the need for intensive manual tuning of hyperparameters. The use of a Gauss-Newton algorithm allows the direct application of the underlying physics of the pose estimation problem, i.e., the perspective equations for quantifying the uncertainty in the estimated pose. In contrast to current learning-based methods, the SPN method can be trained using solely synthetic images of a target spacecraft and is shown to generalize its performance on flight imagery of the same target spacecraft. This research also demonstrates that the SPN method can be used for target-in-target pose estimation to handle terminal stages of a docking scenario where only a partial view of the target spacecraft is available. A unique contribution of this research is the generation of the Spacecraft Pose Estimation Dataset (SPEED), which is used to train and evaluate the performance of pose estimation methods. SPEED consists of synthetic images created by fusing OpenGL-based renderings of a spacecraft 3D model with actual meteorological images of the Earth. SPEED also consists of actual camera images created using a seven degrees-of-freedom robotic arm, which positions and orients a vision-based sensor with respect to a full-scale mock-up of a spacecraft. SPEED is being used to host an international competition on pose estimation in collaboration with the European Space Agency. Infinite Orbits is adopting the pose estimation methods developed during this research for onboard deployment during their commercial on-orbit servicing missions.