Real-time UAV-based geospatial mapping : integrating YOLOv11 for road intersection detection and photogrammetry for 3D terrain modeling on edge devices

Abstract

Uncrewed aerial vehicle (UAV)-based geospatial mapping is pivotal for applications such as urban planning, autonomous navigation, and disaster response, offering high-resolution data collection from inaccessible areas. Traditional methods often rely on expensive sensors or lack real-time processing capabilities, limiting their practicality in resource-constrained environments. This thesis proposes an innovative and cost-effective approach that integrates deep learning and photogrammetry to enable real-time road intersection detection and three-dimensional (3D) terrain modeling using standard color cameras in UAVs, addressing the need for accessible and efficient geospatial solutions.

This research is motivated by the increasing need for precise geospatial information in contexts constrained by limited computational and financial resources. Challenges include achieving robust object detection under diverse environmental conditions, generating precise Digital Elevation Models (DEMs) without ground control points (GCPs), and ensuring real-time processing on edge devices. The methodology leverages a custom dataset of 444 UAV-captured images to train and optimize a YOLOv11m model for road intersection detection, achieving a mean average precision (mAP) of 0.97 and utilizes Structure-from-Motion (SfM) alongside Multi-View Stereo (MVS) techniques to produce DEM from 356 geotagged images with a ground sampling distance (GSD) of 1.5 cm.

The proposed system offers several advantages, including high detection accuracy, detailed terrain modeling, and real-time execution on edge devices like the NVIDIA Jetson Orin Nano, optimized with TensorRT for efficiency. Despite limitations in georeferencing accuracy due to the absence of GCPs, the system achieves reliable detection and fine-resolution terrain mapping, making it a suitable and cost-effective solution for practical deployment. By integrating YOLOv11-based detection with SfM-derived DEMs, this research advances UAV-based mapping, providing a unified open-source framework that lays the foundation for future autonomous systems and supports applications that require terrain-aware geospatial analysis.