Autonomous Navigation around Asteroids using Convolutional Neural Networks

Abstract

Missions to small bodies are increasingly gaining interest as they might hold the secrets to our solar system’s origin while some are also posing a threat to life on Earth. The small size and irregular shape result in complex dynamics complicating the close-proximity operations. Furthermore, due to the long round-trip time communication delays of up to 20 minutes can exists, excluding any required computation time on Earth. Currently used approaches either rely on detecting pre-defined landmarks on the target, detecting features and matching them to a database, or tracking craters or unknown features across images (relative navigation). However, these methods rely heavily on a-priori information, suffer from computationally intensive matching steps, or depend on the accuracy of the initial state estimate.

This work investigated the usage of a novel CNN-based pipeline that can be used to autonomously navigate accurately around asteroids. A top-down CNN-based feature detector is developed consisting of an object and keypoint detection network in sequence, which detects n pre-defined keypoints designated on the target’s 3D model using the 3D SIFT algorithm. These 2D detections alongside their 2D-3D correspondences are sent to an Efficient Perspective-n-Point (EPnP) solver that solves the Perspective-n-Points (PnP) problem. The CNN-based feature detector replaces traditional hand-engineered Image Processing (IP) algorithms as it is more robust to illumination conditions and image noise. Furthermore, the use of a CNN facilitates an offline feature selection step and as such avoid the cumbersome and computationally intensive 2D-3D matching step of the detected 2D feature to their location on the 3D model, plaguing traditional approaches. This pose estimation pipeline can be used to navigate around the asteroid up until it covers the full field of view of the camera, and it can be used to(re)-initialize the navigation filter for a relative navigation approach.

The networks have been selected based on their applicability to embedded devices and this resulted in the use of the SSD-MobileNetV2-FPN-Lite as the object detection network and the Lightweight Pose Network (LPN) model as the keypoint detection network. This lightweight CNN-based pipeline has a fraction of the parameters and Floating Point Operations (FLOPs) compared to state-of-the-art deep-learning networks and pipelines. These networks have been trained and evaluated on synthetic datasets created in this work, consisting of 32,352 images with a variety of poses for a distance of 4.5 km to 9 km from the asteroid for different illumination conditions, asteroid orientations, and image corruptions that emulate real sensor artifacts.

The pipeline could achieve a mean and median line-of-sight distance estimate of around 42 m and 30 m, respectively, at a confidence level of 90% for the large relative range, while satisfying the accuracy requirement of a maximum 10% knowledge error for 99.979% of the cases. Furthermore, the pipeline has been proven to be robust against illumination conditions, occlusions, textures, and image corruptions, mimicking effects of real sensors and the space environment. Demonstrating the efficacy of this CNN-based approach for autonomous navigation around asteroids.