Bringing Visual(-inertial) Odometries to the Limits: In Search of Robust and Accurate State Estimation for Micro Aerial Vehicles in Challenging Mars-Analog Environments

Kurzfassung

In this paper, we present a preliminary performance evaluation of state-of-the-art visual(-inertial) odometries on the INSANE dataset. This dataset contains sequences of an MAV mainly flying through Mars-analog sceneries, equipped with a stereo camera, a monocular camera and an IMU, among other sensors. Challenges arise because of high flying altitudes, fast maneuvers, varying lighting conditions and a homogeneous, unstructured environment. We focus on classical, feature tracking based approaches and evaluate accuracy, robustness and the efficiency of the selected algorithms. With default configurations, VINS-Mono and Granite, a modified version of Basalt with mono-VO capabilities, performed best, while VINS-Fusion and OpenVINS failed to complete the two tested Mars-like sequences