Extraction of building roof countors through integration of high-resolution aerial images and LASER data, using Markov random field

Author(s)

Vanessa Jordão Marcato Fernandes

• Document type: Doctoral Thesis
• Press: Presidente Prudente. 2017-02-03.
• Institution: Universidade Estadual Paulista (Unesp). Faculdade de Ciências e Tecnologia. Presidente Prudente
• Defense date: 2016-12-19
• Advisor: Aluir Porfírio Dal Poz

Abstract

This paper proposes a method for the automatic extraction of building roof contours through a combination of Airborne Laser Scanner (ALS) and photogrammetric data, and Markov Random Field (MRF). Initially, a normalized digital surface model (nDSM) is generated on the basis of the difference between the digital surface model and the digital terrain model, obtained from the LiDAR point cloud. Then the nDSM is segmented to obtain the polygons representing aboveground objects. These polygons are projected onto image to restrict the search space for image segmentation into regions. This process enables the extraction of polygons in the image representing aboveground objects. Building roof contours are identified from among the aboveground objects in the image by optimizing a Markov-random-field-based energy function that embodies roof contour specific properties. In the MRF model are used both polygons extracted from image and from ALS data. The energy function is optimized by the Genetic Algorithm (GA) method. The method proposed in this work was evaluated based on real data - high-resolution aerial images and ALS data. The results obtained in the experimental evaluation showed that the methodology works adequately in the task of extracting the contours of building roofs. The proposed energy function associated with the GA optimization method correctly differentiated the building roof contours from the other high objects present in the scenes. The extracted roof contours show good quality, which is evidenced by the indexes of completeness and correctness obtained by numerical evaluation. Based on the mean indexes obtained for each experiment, the average completeness and correctness for the experiments were equal to 90.96% and 98.99%, respectively. The maximum completeness and correctness values are 99.19% and 99.94%, respectively, and the minimum values are 78.08% and 97.46%, respectively. The lowest values of completeness are associated to the vegetation occlusion areas and presence of shadows. (AU)

• FAPESP's process: 12/22332-2 - Extraction of building roof countors through the integration of high-resolution aerial images and LiDAR data, using Markov random field
• Grantee: Vanessa Jordão Marcato Fernandes
• Support Opportunities: Scholarships in Brazil - Doctorate