ESA GNC Conference Papers Repository
SELECTING KEY FEATURES OF 3D OBJECT MODEL FOR RELATIVE POSE ESTIMATION
The problem which is addressed in this paper refers to the discussion, how complex should be the features extracted from the 3D mesh of a target, in order to get a stable and accurate relative pose estimation with monocular cameras during an On-Orbit Servicing (OOS) mission scenario.The process of the key points extraction from 3D models is very challenging task because of following issues. A) There is no strict definition how the interesting key features should be defined. In most cases it is dependent on the features, which are going to be detected on the space object. Very common features are corners or lines of the spacecraft, which can be detected from the visual data for further matching. Considering these parts, we have to build a 3D key features of the target accordingly. B) The up-to-date target model may be only available during the mission, e.g. reconstruction during inspection phase. When we are still on the ground, we can only have an approximate 3D model of the target. Then the question arises how do we define the model key points on the ground before the OOS mission takes place, if these points could be no more exist on the real space object? It should be defined - if an on-board algorithm is needed or the data should be processed on the ground. C) The limited computing power on board. The processing time for the model key extraction on the ground is not a problem. The challenge is feature processing tasks in space using an On-Board Computer (OBC). Dependent on the platform of OBC, the extraction of 2D image points with it following matching should be fast enough to process live images. We propose two models: the first model contains manually extracted key points; the second one has key points extracted using HarrisKeypoint3D technique from PCL Library. The automatic process of extracting 3D key points with HarrisKeypoint3D function can be used with different 3D meshes of the space object. Thus, can be applied in different OOS missions. We are going to present the results of relative pose estimation using two different models of the target mockup located at European Proximity Operations Simulator (EPOS), German Aerospace Center. Some camera images and ground truth are collected for pose estimation tests also from EPOS facility.