Model-based Pose Estimation of Non-cooperative Aircraft with Zero-shot Machine Learning

Abstract

This thesis explores machine learning methods for the task of drone-pose estimation, with the objective of evaluating the performance of a contemporary method. To facilitate this evaluation, a synthetic dataset is generated, covering a wide range of different conditions. The dataset includes ground truth pose annotations, enabling a quantitative assessment of the method’s accuracy. The model-based zero-shot method SAM-6D was selected for the experiments and evaluated on the synthetic dataset. The results show that SAM-6D is capable of delivering very accurate results under certain conditions. Key factors influencing performance include the complexity of the scene and the distance between the drone and the camera. Additionally, the experiments demonstrated the critical need for an upstream object detection method to enhance the accuracy and reliability of the SAM-6D method. Furthermore, a qualitative analysis of the results obtained by SAM-6D on real data without pose annotations was conducted to confirm the plausibility of the previous results.