New capabilities toward trustable pose estimation using AI/ML for Satellite-to-Satellite imagery is introduced. This approach applies The Aerospace Corporation’s patent pending method called BetterNet to imagery data from the ExoRomper in-space AI/ML Pose Estimation testbed aboard the Slingshot-1 SmallSat mission. BetterNet introduces the ability to algorithmically identify individual out-of-distribution predictions of an AI model during operational use based on an assessment of local training support in the pose estimation algorithm’s AI stages through statistical modeling of activation data from an AI model’s hidden layers as well as introduce the use of this capability as a diagnostic for assessing and explaining poor AI model performance or to select from candidate AI models for operational deployment. Additionally, BetterNet produces human-interpretable performance metrics beyond the AI model outputs intended during training based on analysis and modeling of features in latent space representations of the model.For the application of satellite-to-satellite imagery presented, BetterNet is applied to a pose estimation pipeline consisting of an AI stage for object detection, an AI stage for keypoint detection, and a numerical optimization stage using Perspective-n-Point with RANSAC for final pose estimation. Imagery data was sourced from synthetic, laboratory, and the real ExoRomper inspace testbed to train ExoRomper’s pose estimation. Candidate models for the AI stages of the ExoRomper testbed were drawn from open-source development and included YOL-Ov3 Tiny or YOLOv5 nano for object detection and MSPN or MobileNetV3Large for keypoint detection.The ability of BetterNet to exploit features unintended in an AI/ML model’s training is demonstrated by prediction of attitude error from ExoRomper’s keypoint detection stage. When the AI model trained for the sole intended purpose of detecting keypoints of a miniature satellite mockup was analyzed using BetterNet, the features learned for keypoint detection indicate correlation to the resulting attitude error in pose estimation and may be used to provide augmented confidence for the pose estimation algorithm’s results. These combined capabilities enable per-input confidence scores for AI model predictions, which may be used for operational monitoring and assessment of model predictions either to signal review by human satellite operators or enable autonomous decision-making.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    Capabilities Toward Trustable AI/ML Pose Estimation for Satellite-to-Satellite Imagery


    Contributors:


    Publication date :

    2024-03-02


    Size :

    3490053 byte




    Type of media :

    Conference paper


    Type of material :

    Electronic Resource


    Language :

    English



    Making Trustable Satellite Experiments: An Application to a VoIP Scenario

    Auger, Antoine / Lochin, Emmanuel / Kuhn, Nicolas | IEEE | 2019


    Satellite Pose Estimation Using Modified Residual Networks

    Uma Rani, M. / Thangavel, Senthil Kumar / Lagisetty, Ravi Kumar | Springer Verlag | 2022


    Trustable Electronics - TRUST-E

    Wahl, Michael / Stich, Peter / Busch, Rebecca | VDE-Verlag | 2022


    Revisiting Monocular Satellite Pose Estimation With Transformer

    Wang, Zi / Zhang, Zhuo / Sun, Xiaoliang et al. | IEEE | 2022


    Commercial Satellite Imagery in Peacekeeping Sensor Capabilities - Present and Near Future

    Paquette, J.-P. | British Library Conference Proceedings | 2004