Identifying obstacles and key measurements of roof surfaces using a digital surface model and an orthomosaic
(2021) In Master's Theses in Mathematical Sciences FMAM05 20211Mathematics (Faculty of Engineering)
- Abstract
- This thesis proposes a machine-learning-supported pipeline which, put together, provides an estimate of roof surface area for solar panel installation. This pipeline is comprised of roof segmentation based on images and digital surface models and provides subsequent identification of key measurements such as roof length, roof width and roof angle. Obstacles belonging to a roof are also identified and considered in the final estimation.
The pipeline consists of a Faster RCNN model built with Detectron2 to help in solving segmentation issues and a roof segmentation algorithm that utilizes a contour-following algorithm and the Ramer-Douglas-Peucker algorithm which approximates polygons by removing points which does not contribute enough to... (More) - This thesis proposes a machine-learning-supported pipeline which, put together, provides an estimate of roof surface area for solar panel installation. This pipeline is comprised of roof segmentation based on images and digital surface models and provides subsequent identification of key measurements such as roof length, roof width and roof angle. Obstacles belonging to a roof are also identified and considered in the final estimation.
The pipeline consists of a Faster RCNN model built with Detectron2 to help in solving segmentation issues and a roof segmentation algorithm that utilizes a contour-following algorithm and the Ramer-Douglas-Peucker algorithm which approximates polygons by removing points which does not contribute enough to the shape. The algorithm concurrently classifies roof components as roof ridge, roof gutter and the sides of the roofs.
The results of the pipeline are then evaluated as separate components and show varying results for building segmentation, measuring the roof and detecting obstacles. The results for the building segmentation and obstacle detection are evaluated using mean average precision at different intersection over union bounds and show acceptable results. The roof segmentation results are evaluated using the Euclidean distance and show good precision.
A major bottleneck for this project is the number of available images. It is hypothesized that better results could be obtained through having a larger dataset, improving each algorithm in the pipeline to be able to support more roof types and adding a roof classification step to the pipeline. (Less) - Popular Abstract
- To automate the design of solar panel installation layouts, this work presents a pipeline which automates the measuring and obstacle detection of roofs using drone imagery and a height data of the image surface. A solar panel layout is the grid in which the solar panels are installed on the roof surface and is usually placed on available roof surface with a certain distance from the roof edges, avoiding areas with e.g. chimneys, ladders and windows.
The pipeline which is presented in this work, consists of a machine learning model that is trained on high resolution drone imagery to identify and localize objects such as buildings, roofs and obstacles, and an algorithm which refines the localization slightly, separates the roofs of the... (More) - To automate the design of solar panel installation layouts, this work presents a pipeline which automates the measuring and obstacle detection of roofs using drone imagery and a height data of the image surface. A solar panel layout is the grid in which the solar panels are installed on the roof surface and is usually placed on available roof surface with a certain distance from the roof edges, avoiding areas with e.g. chimneys, ladders and windows.
The pipeline which is presented in this work, consists of a machine learning model that is trained on high resolution drone imagery to identify and localize objects such as buildings, roofs and obstacles, and an algorithm which refines the localization slightly, separates the roofs of the building and lastly identifies important components of the roofs.
Once all of the necessary components have been identified, the components are measured using the height data and the geographical coordinates (longitude and latitude) which are encoded in the drone imagery. The measurements obtained by this pipeline include the angle of the roof, the width of the roof and the length of the roof.
The results show decimeter precision and an error within 5 degrees of the angle accuracy in all of the measured buildings. This provides a good starting point for further improvements of the pipeline. These improvements are believed to be reachable by using a larger training dataset when training the machine learning model and by improving the algorithm further to both include a larger quantity of roof types. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9054785
- author
- Schön, Maximilian LU
- supervisor
- organization
- alternative title
- Identifiering av hinder och viktiga mätvärden från takytor genom användning av digitala höjdmodeller och ortofoto
- course
- FMAM05 20211
- year
- 2021
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- roofs, image segmentation, image analysis, digital surface model, machine learning, orthomosaic, software pipeline
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUTFMA-3444-2021
- ISSN
- 1404-6342
- other publication id
- 2021:E25
- language
- English
- id
- 9054785
- date added to LUP
- 2021-06-28 14:21:50
- date last changed
- 2021-06-28 14:21:50
@misc{9054785,
abstract = {{This thesis proposes a machine-learning-supported pipeline which, put together, provides an estimate of roof surface area for solar panel installation. This pipeline is comprised of roof segmentation based on images and digital surface models and provides subsequent identification of key measurements such as roof length, roof width and roof angle. Obstacles belonging to a roof are also identified and considered in the final estimation.
The pipeline consists of a Faster RCNN model built with Detectron2 to help in solving segmentation issues and a roof segmentation algorithm that utilizes a contour-following algorithm and the Ramer-Douglas-Peucker algorithm which approximates polygons by removing points which does not contribute enough to the shape. The algorithm concurrently classifies roof components as roof ridge, roof gutter and the sides of the roofs.
The results of the pipeline are then evaluated as separate components and show varying results for building segmentation, measuring the roof and detecting obstacles. The results for the building segmentation and obstacle detection are evaluated using mean average precision at different intersection over union bounds and show acceptable results. The roof segmentation results are evaluated using the Euclidean distance and show good precision.
A major bottleneck for this project is the number of available images. It is hypothesized that better results could be obtained through having a larger dataset, improving each algorithm in the pipeline to be able to support more roof types and adding a roof classification step to the pipeline.}},
author = {{Schön, Maximilian}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Identifying obstacles and key measurements of roof surfaces using a digital surface model and an orthomosaic}},
year = {{2021}},
}