Drones Autonomous Landing Scene Detection with Transfer Learning

Abstract

This paper proposes an improved method for autonomous landing scene detection for drones. The study addresses challenges that arise when similar environments appear different at varying altitudes. Using deep learning methods and a hybrid ensemble technique, the proposed system improves the accuracy and reliability of landing scene recognition. The proposed system achieved approximately 97.65% accuracy using transfer learning models such as ResNet50 and ResNext50 combined with a hybrid Random Forest classifier. Transfer learning techniques using ResNet50 and ResNeXt50 models are applied to the LandingScenes-7 dataset to identify safe landing locations in real time. Thresholding techniques and a novelty detection module enable the system to handle unpredictable environmental conditions and provide confidence-based classification decisions. This research has significant applications in drone technology, particularly in logistics, emergency response, and surveillance. The proposed system enhances drone intelligence and improves operational safety in dynamic environments by enabling reliable autonomous landing decisions.