Deep neural networks have had profound significance in addressing visual object detection and classification tasks. However, though with the caveat of needing large amounts of annotated training data. Furthermore, the possibility of neural networks overfitting to the biases and faults included in their respective datasets. In this work, methods for achieving robust neural networks, able to tolerate untrusted and possibly erroneous training data, are explored. The proposed method is shown to improve performance and help neural networks learn from untrusted data, provided a thoroughly annotated subset.

Autonomous ships rely on sensory data to perceive objects of interest in their environment. Deep Learning based object detection in the image domain commonly used to solve this issue. The robustness of such approaches in non-ideal conditions is, however, still to be proven. In this work state of the art methods are applied on the RetinaNet architecture attempting to create a more robust object detection network given noisy input data. The GroupSort activation function and Spectral Normalization is used and the results are compared to the standard RetinaNet network. Our findings show that these modifications perform better and ensure robustness under moderate noise levels, than the standard RetinaNet network.

No abstract

No abstract

In this paper, we present an end-to-end simulation framework for tracking an uncooperative Target spacecraft in Low Earth Orbit using a CubeSat-class Ego spacecraft outfitted with a camera. Currently, capturing high-fidelity realistic images in space for this scenario is difficult and exorbitantly expensive. Therefore, we developed a framework to simulate the spacecraft orbits in Basilisk software and generate high-fidelity realistic images of spacecraft in Unreal Engine, including the effects from Sun, Earth, Moon and stars.The Ego spacecraft uses cameras to capture images of the uncooperative Target and estimates its position and attitude using a CNN based 6DOF pose estimation pipeline, eliminating need for large SWAP-C(Size, Weight, Power and Cost) sensors like LI-DAR or reliance on inter-spacecraft communication, This CNN, which is motivated by ESA's Pose Estimation challenge of 2019, is trained using simulated data from our end-to-end simulation framework. We compare the performance of two distinct CNNbased algorithms for pose estimation along a nominal trajectory. In presence of non-Gaussian modeling uncertainties, the statedependent estimation error is characterized with a quadratic upper-bound. The quadratically-bounded error can be used by a robust controller to maneuver Ego spacecraft to track the uncooperative Target.