Event-Based Detection and 3-DoF Pose Estimation of a UAV Docking Platform

Sammanfattning

This thesis investigates an event-based perception pipeline for autonomous drone landing using data derived from RGB-to-event conversion. The system combines deep learning models for two core tasks: (1) classification of whether a target drone is within the field of view, and (2) regression of the relative 3D position of a docking station. Both models are trained primarily in simulation and evaluated in both simulated and real-world environments to assess their generalization capability under domain shift. The classification model achieves strong performance in simulation, with an F1-score of 94.72%, but shows a significant degradation in real-world deployment, reaching 67.99% on the full dataset and improving to 78.57% after filtering sparse-event samples. The regression model demonstrates moderate accuracy in simulation, with a mean 3D distance error of 1.14 m, but exhibits increased instability and reduced consistency in real-world scenarios, particularly in depth estimation. The results highlight a clear sim-to-real performance gap, primarily driven by differences in event data characteristics, motion blur in RGB-to-event conversion, and limited dataset diversity. Additionally, temporal sparsity and dataset imbalance were identified as key factors affecting robustness in both tasks. Overall, the study demonstrates that event-based perception using simulated event generation is a viable approach for autonomous navigation tasks, but also emphasizes the need for improved data collection strategies, more representative event modeling, and stronger temporal robustness for real-world deployment.