Extracting vegetation to generate realistic natural scenes for virtual reality: from drone images to Unreal Engine
(2025) In Student thesis series INES NGEM01 20251Dept of Physical Geography and Ecosystem Science
- Abstract
- Realistic virtual environments offer valuable applications in rehabilitation, especially for individuals with limited mobility. However, most existing virtual reality (VR)-based rehabilitation platforms rely on procedurally generated or low-detail natural scenes, often lacking ecological and spatial authenticity. This study proposes a complete and cost-effective workflow that combines unmanned aerial vehicle (UAV)-based Structure-from-Motion (SfM) photogrammetry with real-time 3D rendering in Unreal Engine 5 (UE5) to generate an ecologically grounded virtual scene.
The virtual scene represents a section of Klosterängshöjden Park in Lund, specifically its cycling path, which is surrounded by groups of trees and bushes. To measure the... (More) - Realistic virtual environments offer valuable applications in rehabilitation, especially for individuals with limited mobility. However, most existing virtual reality (VR)-based rehabilitation platforms rely on procedurally generated or low-detail natural scenes, often lacking ecological and spatial authenticity. This study proposes a complete and cost-effective workflow that combines unmanned aerial vehicle (UAV)-based Structure-from-Motion (SfM) photogrammetry with real-time 3D rendering in Unreal Engine 5 (UE5) to generate an ecologically grounded virtual scene.
The virtual scene represents a section of Klosterängshöjden Park in Lund, specifically its cycling path, which is surrounded by groups of trees and bushes. To measure the vegetation cover, three types of SfM-derived data products were employed: dense point cloud, photogrammetric 3D mesh, and high-resolution digital orthophoto map (DOM). These were preprocessed using statistical noise filtering, mesh vertex extraction, and tiled raster techniques. Vegetation segmentation and height estimation were conducted using methods suited to each data type. For point clouds and mesh vertices, tree detection was carried out in ENVI LiDAR by applying a series of minimum height thresholds to evaluate segmentation performance. For orthophotos, tree crowns were segmented using the Segment Anything Model (SAM), and height was estimated via Digital Surface Model (DSM)-Digital Elevation Model (DEM) raster differentiation. Ground-truth data, manually annotated from the original point cloud and supported by limited field measurements, served as the reference for accuracy evaluation.
Among the three data types, the point cloud-based workflow achieved the highest performance, yielding an F1-score of 70.79% for ddetection and an R² of 0.92 for tree height estimation. In contrast, mesh- and DOM-based methods were less precise in structure detection but proved effective for enhancing visual realism in the rendered environment. Vegetation attributes, including location and height, were integrated into UE5 using a data-driven placement system built with Blueprint scripting. Predefined high-resolution vegetation models were selected and scaled accordingly to match field conditions. Rendering optimizations such as Level of Detail (LOD) management, Nanite geometry, and instancing techniques were also employed to ensure real-time performance and visual quality.
This research shows that UAV-SfM data can support the creation of virtual environments that are both rooted in real vegetation and visually impressive. The workflow suggested in the research is practical, scalable, and suitable for applications in environmental visualization, rehabilitation support, and interactive digital twin systems. (Less) - Popular Abstract
- Being confined indoors for extended periods can be challenging, especially for individuals recovering from illnesses or injuries that restrict movement. Research has shown that experiencing nature can significantly support the healing process. This project offers a creative solution by allowing people to enjoy realistic outdoor scenes through virtual reality (VR) environments.
To start, I used a drone to capture hundreds of aerial photographs of a park in Lund, Sweden. These images were processed into 3D models using specialized software, enabling a detailed representation of the trees, bushes, and terrain. Advanced techniques were then used to identify the exact locations and heights of the trees. All this information was then brought... (More) - Being confined indoors for extended periods can be challenging, especially for individuals recovering from illnesses or injuries that restrict movement. Research has shown that experiencing nature can significantly support the healing process. This project offers a creative solution by allowing people to enjoy realistic outdoor scenes through virtual reality (VR) environments.
To start, I used a drone to capture hundreds of aerial photographs of a park in Lund, Sweden. These images were processed into 3D models using specialized software, enabling a detailed representation of the trees, bushes, and terrain. Advanced techniques were then used to identify the exact locations and heights of the trees. All this information was then brought into a powerful software tool called Unreal Engine 5, which is commonly used to build video games. However, in this project, it was used to replicate a real wild bike track in a virtual setting.
Instead of relying on simple or generic tree models, I placed each tree in its correct real-world location and assigned it the right size. I used smart tools to ensure the virtual forest runs smoothly on computers. For example, faraway trees are displayed with simpler shapes, and only the trees near the user are shown in full detail.
The result is a beautiful and lifelike digital grassland with trees and bushes that match the actual park. This environment can be used in therapy to give patients the experience of being in nature without having to leave their homes. It could also help urban planners, environmental scientists, or anyone who needs a detailed digital copy of a natural space. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9211388
- author
- Zhang, Xiaoyue LU
- supervisor
-
- Per-Ola Olsson LU
- Anna Terskaia LU
- organization
- course
- NGEM01 20251
- year
- 2025
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Physical Geography and Ecosystem analysis, UAV photogrammetry, 3D vegetation modeling, virtual reality, Unreal Engine 5
- publication/series
- Student thesis series INES
- report number
- 717
- language
- English
- id
- 9211388
- date added to LUP
- 2025-09-02 09:42:45
- date last changed
- 2025-09-02 09:42:45
@misc{9211388,
abstract = {{Realistic virtual environments offer valuable applications in rehabilitation, especially for individuals with limited mobility. However, most existing virtual reality (VR)-based rehabilitation platforms rely on procedurally generated or low-detail natural scenes, often lacking ecological and spatial authenticity. This study proposes a complete and cost-effective workflow that combines unmanned aerial vehicle (UAV)-based Structure-from-Motion (SfM) photogrammetry with real-time 3D rendering in Unreal Engine 5 (UE5) to generate an ecologically grounded virtual scene.
The virtual scene represents a section of Klosterängshöjden Park in Lund, specifically its cycling path, which is surrounded by groups of trees and bushes. To measure the vegetation cover, three types of SfM-derived data products were employed: dense point cloud, photogrammetric 3D mesh, and high-resolution digital orthophoto map (DOM). These were preprocessed using statistical noise filtering, mesh vertex extraction, and tiled raster techniques. Vegetation segmentation and height estimation were conducted using methods suited to each data type. For point clouds and mesh vertices, tree detection was carried out in ENVI LiDAR by applying a series of minimum height thresholds to evaluate segmentation performance. For orthophotos, tree crowns were segmented using the Segment Anything Model (SAM), and height was estimated via Digital Surface Model (DSM)-Digital Elevation Model (DEM) raster differentiation. Ground-truth data, manually annotated from the original point cloud and supported by limited field measurements, served as the reference for accuracy evaluation.
Among the three data types, the point cloud-based workflow achieved the highest performance, yielding an F1-score of 70.79% for ddetection and an R² of 0.92 for tree height estimation. In contrast, mesh- and DOM-based methods were less precise in structure detection but proved effective for enhancing visual realism in the rendered environment. Vegetation attributes, including location and height, were integrated into UE5 using a data-driven placement system built with Blueprint scripting. Predefined high-resolution vegetation models were selected and scaled accordingly to match field conditions. Rendering optimizations such as Level of Detail (LOD) management, Nanite geometry, and instancing techniques were also employed to ensure real-time performance and visual quality.
This research shows that UAV-SfM data can support the creation of virtual environments that are both rooted in real vegetation and visually impressive. The workflow suggested in the research is practical, scalable, and suitable for applications in environmental visualization, rehabilitation support, and interactive digital twin systems.}},
author = {{Zhang, Xiaoyue}},
language = {{eng}},
note = {{Student Paper}},
series = {{Student thesis series INES}},
title = {{Extracting vegetation to generate realistic natural scenes for virtual reality: from drone images to Unreal Engine}},
year = {{2025}},
}