With new regulations around unmanned aerial vehicles (UAV) use in the European Union, the drone industry is set to bloom. This led to theBEAST-project, which aims at the automatic fail-safe landing for UAVs. This technical paper tackles a specific part within the module, which revolves around anomaly detection to avoid obstacles for fail-safe landing, where the CutPaste anomaly detection framework is used. For this framework new alterations are introduced, CutPaste-multiple and CutPaste-normal scar. Various experiments try to maximize the model’s performance, including choosing the tile size, CutPaste alteration method, hyperparameter tuning, and different data augmentation techniques. The best-performing model achieved an F1-score of 0.8036. Although anomaly detection looks like a suitable approach to resolve this problem, some limitations must be overcome before being used in a real-life application.
Click here for the poster of this project.