Lidar-based 6D Pose Estimation of a Satellite in an On-Orbit-Servicing Simulator (OOS-SIM) Scenario

Abstract

Since the invention in early of the 1960s, Light Detection And Ranging (LiDAR)s have been used for decades as a sensor for observing objects and perceiving surrounding environment from afar. Due to high density and great accuracy data that is being produced from the scans, LiDAR sensors are highly suitable to be applied in autonomous robotic applications. Specifically at DLR, a system of LiDARs is being implemented in the On-Orbit-Servicing Simulator. 3D data offers the advantage of accurate long-range detection and tracking of satellites in space, enabled by novel space-qualified LiDAR systems. Point cloud based satellite detection and tracking methods can be developed, compared, and tested with DLR's realistic satellite mockup in the On-Orbit-Servicing Simulator robotic setup. While plenty of traditional computer vision techniques existed and were proven with good track of record on 6D pose estimation problem, the suitability for long-range detection on space-qualified applications with Deep Learning based methods for depth-based 6D pose estimation using LiDARs data needs to be analyzed. For that work, a prior extrinsic calibration of the LiDARs to each other is needed to be implemented. This thesis work is focusing on two main components: Firstly, a setup of a dual-LiDAR modality is being implemented for point cloud perception of the satellite. Prior to further processing actions, the two LiDARs are being calibrated such that the point cloud data perceived from both LiDARs is consistently and accurately aligned. Accordingly, meaningful data can be used for later processing steps. Secondly, the data perceived from the two LiDARs is then being utilized for the 6D pose estimation of the satellite. The 6D pose estimation of the satellite with LiDARs point cloud data is considered and evaluated at far range distance.