Object-Based Building Boundary Extraction from Lidar Data You Shao and Samsung Lim

Most filtering algorithms require rasterisation of lidar data Additional computing overhead Loss of information Increase of uncertainty Our method No rasterisation Adaptive window size Morphological filtering DTM generation and building detection Research Objectives

The UNSW Campus (1 km x 2 km) Small residential buildings, high-rise buildings, steep roads, tall trees and large green areas Lidar data (X, Y, Z, I) Airborne imagery (R, G, B) 2-year gap between the two datasets Study Area and Datasets

Aerial Ortho-photo

Lidar Intensity

Vertical Profile

Employ dilation and erosion to find the maximum or minimum measurements in lidar points An adaptive window size indicator is developed to detect building rooftops and modify the window size automatically An approximate size of a building can be detected by measuring the elevation rise and fall, and therefore the window size can be changed accordingly Proposed Adaptive Filtering

Adaptive Filtering (Workflow)

Normalised Difference Vegetation Index (NDVI) to remove vegetation Alpha-shape to form building outlines Grid-based algorithm Modified convex hull algorithm Fine-tuning with adjustable parameters to remove small residuals Approaches to Building Detection

Extracted Buildings

Unfiltered Classification Results in Residential Area

Filtered Classification Results in Residential Area

Accuracy Assessment

Alpha-shape algorithm

Grid-based algorithm

Modified convex hull algorithm

Alpha-shape Modified convex hull Boundary Extraction (1/2)

Grid-based Boundary Extraction (2/2)

B1B2B3B4B5B6B7B8Mean Alpha- shape Modified convex hull Grid- based Horizontal RMSE (m, 1σ)

Building extraction in residential areas (Site 1)

Building extraction in residential areas (Site 2)

Building extraction accuracy

The proposed algorithm is suitable for steep urban areas with varying building sizes The required parameters of the proposed algorithm can be automatically determined The test results show that the proposed algorithm is able to classify ground points with a vertical accuracy of 36 cm, a horizontal accuracy of 75 cm and a commission error less than 6% As for multi-rooftop buildings, it is difficult to determine the actual size of the building; however, this problem can be solved by the proposed dual-direction process Concluding Remarks