Search

This paper investigates the use of convolutional neural networks for the purpose of image foreground extraction from a dynamic environment. The proposed solution utilises the latest developments in image segmentation using pixel-wise classification to produce foreground target extraction for real-time operations. A collection of spacecraft images were assembled for network training and evaluation. The proposed technique takes advantage of transfer learning for the stable training of a convolutional neural network classifier. The image extraction software was applied to a thermal camera video, taken by an undocking spacecraft from the International Space Station. The results show the proposed deep learning-based image extraction has advantages over traditional background subtraction methods. This investigation provides evidence that semantic segmentation using convolutional neural network can be an effective tool for spacecraft image isolation and extraction from a dynamically cluttered scene.

This paper proposes several enhancements to the softPOSIT algorithm with applications to spacecraft pose estimation using a monocular camera. First, the proposed enhancements include a technique for reducing false matches as result of local minimum trapping. Second, this paper provides two strategies for iteration control parameter initialization by using the trace of the correspondence distance, and by using image centroid matching. The method of image centroid matching allows the world model center of geometry to align with the image centroid. The alignment result in reasonable correspondence weighting values used for match optimization. The various algorithm enhancements were tested on 26,180 simulations with varying geometries and initial pose conditions. Results show a significant increase in accuracy when compared with the original method.

This paper presents the design of a pose estimation method for autonomous robotic proximity operations with an uncooperative target using a single infrared camera and a simple three-dimensional model of the target. Specically, the presented method makes use of the so-called SoftPOSIT algorithm to solve the model-to-image registration problem. This particular method is found to be most useful for ranges where surface features are not well resolved, that is, from approximately 30 to 500 meters. The proposed solution is validated in preliminary numerical simulations using low-resolution CAD models of Envisat and ISS that emulate IR images, and obtained results are reported and discussed.

In this paper, a nonlinear direct adaptive output control methodology is developed to address the problem of spacecraft attitude control under large dynamics uncertainties. The main advantage of the proposed approach over other adaptive approaches for spacecraft attitude control is that it can handle significantly large inertia uncertainties, without requiring any on-line estimation of the unknown moments of inertia. Furthermore, the implementation of the controller does not require a mathematical model of the plant as the control gain adaptation mechanism relies on feedforward signals from an ideal model designed to provide a satisfactory response to the desired attitude commands, as well as on output tracking errors between the uncertain spacecraft and the ideal model. By modeling the spacecraft as a square nonlinear state-space plant through the use of the modified Rodrigues parameters allows the system to satisfy the almost striclty passive conditions, which are required to establish the formal proof of stability. The performance of the new adaptive attitude control approach is illustrated in numerical simulations for both a simple rigid-body rest-to-rest maneuver and a high-fidelity ISS free-flyer robotic capture maneuver.