Enhanced Full Visual Slam for Close Proximity Exploration of Asteroids Combining Incremental Smoothing with Filtering Techniques

This study proposes a visual SLAM algorithm for autonomous asteroid exploration, leveraging state-of-the-art estimation techniques from Robotics. Two variants of a factor graphbased SLAM system are developed: ISAM2dynamics, which directly incorporates orbital dynamics constraint into the SLAM estimation process, and ISAM2concurrent, featuring a dynamic filter tightly-coupled with the smoother. Both methods provide accurate sparse metric maps of the target while estimating the entire history of attitude and orbital states. A keyframe insertion logic is also proposed to insert graph states in an adaptive manner and only as required by the algorithm. A medium fidelity image processing emulator is presented to avoid the burden of rendering images while retaining representativeness of the measurements’ errors. This enabled a comparative Monte Carlo analysis to highlight advantages and drawbacks of the approaches. Results show that incorporating the dynamics directly into the factor graph allows a more accurate and consistent estimation with respect to ISAM2concurrent approach. Moreover, it enables the smoother to correct estimation of landmarks that are no longer visible, thus maintaining the entire map consistent throughout the navigation window. However, this comes at the cost of a computational cost of the smoothing that increases over time.