Publication:
A unified optimization framework for hybrid UAV precision landing: Attention-weighted you-only-look once, and region-of-interest-gated class-aware vision.

Loading...
Thumbnail Image
Date
2026-08
Authors
Muhammad Uzair Gill
Journal Title
Journal ISSN
Volume Title
Publisher
Research Projects
Organizational Units
Journal Issue
Abstract
Precision landing in cluttered near-ground environments is essential for autonomous UAV pick-and-drop missions. However, small obstacles within the landing zone are often missed by real-time detectors, and improvements reported in offline evaluations are rarely shown to support reliable landing decisions on embedded hardware. This thesis develops and validates a perception-to-decision framework to address these challenges. During offline model development, YOLO11n is improved using a budgeted hybrid hyperparameter optimisation strategy named Differential Evolution–Multi-Fidelity Optimisation–Bayesian Optimisation (DE-MFO-BO) together with an architectural adaptation named AWAGS-YOLO. AWAGS-YOLO incorporates attention-augmented backbone blocks, a weighted Bi-directional Feature Pyramid Network (BiFPN) neck, Convolutional Block Attention Module (CBAM), Global Attention Mechanism (GAM), and the Scylla Intersection over Union (SIoU) regression loss. The individual and combined effects of these improvements are evaluated under a common protocol on VisDrone2019 and Realm, a 12-class aerial dataset containing 7,631 images introduced in this thesis. Performance is also reported for the COCO small-object category with an area below 32² pixels. Hyperparameter optimisation alone improved mAP(50-95) by 20.96% on VisDrone2019 and 23.29% on Realm compared with the baseline YOLO11n configuration. The combined model achieved mAP50 of 0.773 and mAP(50-95) of 0.457 on Realm. During online validation, detections are converted into descent authorisation using a class-aware region-of-interest (ROI) gate with temporal filtering implemented within a finite-state controller. The complete framework was validated in a closed-loop AirSim mission, achieving a landing error of 0.99 m with no unsafe descent authorisations. The framework was also deployed on a Raspberry Pi 5, where it sustained an end-to-end processing rate of 12.6 frames per second for approximately 10 minutes before thermal throttling occurred. A marker-based module improved final alignment during hover and early descent. However, marker detection dropped to zero below 2 m altitude when the marker left the camera field of view. This limitation is identified in the thesis and discussed in the future work. The results demonstrate that dataset-aware hyperparameter optimisation and architectural adaptation provide complementary improvements in detection performance. They also show that explicit safety-oriented decision logic is required to translate detection quality into reliable autonomous landing behaviour.
Description
Keywords
Citation