Scholarlink

<p>The increasing demand for autonomous drone delivery systems faces a critical challenge: safe and precise landing in complex, dynamic environments. Drones have been identified as a potential solution through rapid delivery and the ability to reach areas inaccessible to conventional vehicles. Despite this, safe navigation and precise landing remain challenges in dynamic environments, making real-time obstacle detection and safe landing zone identification necessary. In this study, deep learning and computer vision integrating a YOLOv8-based model for real-time obstacle detection such as persons, vehicles, trees etc., with a Vision Transformer (ViT) model for precise drone landing space segmentation from aerial images. The YOLOv8 detection rate was evaluated using mean Average Precision (mAP) measures, achieving a high accuracy of 95% at an IoU threshold of 0.5 (mAP@0.5), demonstrating its effectiveness in reliable obstacle detection. This study also uses different transformer-based segmentation models like SegFormer, SAM (Segment Anything Model) and CNN based Mask R-CNN. The study evaluates the models performance based on metrics such as Intersection over Union (IoU) and Dice Score. The study emphasizes that SAM performed best along with YOLOv8 with an IoU and Dice Score of 0.9969, providing accurate segmentation and resilience. In this context the “term resilience” refers to the robustness. This robustness is demonstrated through its ability to handle environmental complexity by cleanly isolating objects within a cluttered scene with varied textures and dynamic elements. It also shows a strong capacity for managing diverse lighting conditions, such as shadows and glare, without a loss in segmentation accuracy. Furthermore, SAM performs with consistency across variations in object scale and position, reliably segmenting objects at different distances, sizes, and orientations. MaskRCNN model followed with 0.7509 and Segformer performed low with 0.0370 mean IoU and Dice score respectively. Lack of a combined system for both obstacle detection and landing zone identification is the biggest limitation. This study fills that gap by merging detection and segmentation to facilitate real-time autonomous landing. First up, object detection is applied for recognizing the obstacles (people, vehicles, or equipment) in the area. Once those are identified, the system switches gears to semantic segmentation for the remaining spaces. That part maps out clear zones for landing or safe navigation routes. In general, transformer-based models, particularly SAM, perform better than others in aerial image segmentation and are thus extremely effective at improving safety, scalability, and efficiency in drone-based delivery systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transformer-Guided Vision for Autonomous Drone Delivery: Integrating YOLOv8 and ViT for Obstacle Detection and Landing Zone Segmentation

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Sourav Mondal,
  • Kamakhya Paul,
  • Sana Sultana Pathan,
  • Gopi Chand Perala,
  • Venkata Karthik Dhulipudi

摘要

The increasing demand for autonomous drone delivery systems faces a critical challenge: safe and precise landing in complex, dynamic environments. Drones have been identified as a potential solution through rapid delivery and the ability to reach areas inaccessible to conventional vehicles. Despite this, safe navigation and precise landing remain challenges in dynamic environments, making real-time obstacle detection and safe landing zone identification necessary. In this study, deep learning and computer vision integrating a YOLOv8-based model for real-time obstacle detection such as persons, vehicles, trees etc., with a Vision Transformer (ViT) model for precise drone landing space segmentation from aerial images. The YOLOv8 detection rate was evaluated using mean Average Precision (mAP) measures, achieving a high accuracy of 95% at an IoU threshold of 0.5 (mAP@0.5), demonstrating its effectiveness in reliable obstacle detection. This study also uses different transformer-based segmentation models like SegFormer, SAM (Segment Anything Model) and CNN based Mask R-CNN. The study evaluates the models performance based on metrics such as Intersection over Union (IoU) and Dice Score. The study emphasizes that SAM performed best along with YOLOv8 with an IoU and Dice Score of 0.9969, providing accurate segmentation and resilience. In this context the “term resilience” refers to the robustness. This robustness is demonstrated through its ability to handle environmental complexity by cleanly isolating objects within a cluttered scene with varied textures and dynamic elements. It also shows a strong capacity for managing diverse lighting conditions, such as shadows and glare, without a loss in segmentation accuracy. Furthermore, SAM performs with consistency across variations in object scale and position, reliably segmenting objects at different distances, sizes, and orientations. MaskRCNN model followed with 0.7509 and Segformer performed low with 0.0370 mean IoU and Dice score respectively. Lack of a combined system for both obstacle detection and landing zone identification is the biggest limitation. This study fills that gap by merging detection and segmentation to facilitate real-time autonomous landing. First up, object detection is applied for recognizing the obstacles (people, vehicles, or equipment) in the area. Once those are identified, the system switches gears to semantic segmentation for the remaining spaces. That part maps out clear zones for landing or safe navigation routes. In general, transformer-based models, particularly SAM, perform better than others in aerial image segmentation and are thus extremely effective at improving safety, scalability, and efficiency in drone-based delivery systems.