Farzaneh Aghighi; Omid Mahdi Ebadati E.; Hossein Aghighi
Abstract
Lidar point cloud dataset and 3-D models are widely used in urban feature extraction, forest, urban and tourism management, robotics, computer game production etcetera. On the other hand, The existence of outliers in the lidar point cloud is inevitable. Therefore, outlier detection and removing them ...
Read More
Lidar point cloud dataset and 3-D models are widely used in urban feature extraction, forest, urban and tourism management, robotics, computer game production etcetera. On the other hand, The existence of outliers in the lidar point cloud is inevitable. Therefore, outlier detection and removing them from lidar point cloud data have been known as necessary steps in lidar point cloud processing. Over the past decade, several outlier detection techniques have been introduced in the literature; however, most of them are time-consuming, expensive, and computationally complicated. For overcoming these limitations, this article introduces a new automatic approach for outlier detection using a support vector machine-based conditional random field (SVM-CRF) technique and box plots methods. In this approach, a box plot analyzes the output energyvector of SVM-CRF to recognize outliers. The methods were evaluated using ISPRS benchmark datasets of Vaihingen provided in order to urban classification and 3D building reconstruction. To evaluate this method, first of all, outliers, that are almost closed to objects, were added to the data set manually. Then the research steps were done to evaluate the proposed method's ability for detecting outliers. The evaluation of this research showed an overall accuracy of 62% as the performance of the proposed model. Although the RANSAC algorithm has better performanc, it is a more costly and time-consuming technique than the proposed outlier detection technique.
F Aghighi; O.M Ebadati; H Aghighi
Volume 9, Issue 2 , December 2017, , Pages 41-60
Abstract
Light Detection and Ranging (LiDAR) point cloud dataset and 3 dimensional (3-D) models have been extensively used for urban feature extraction, urban management, forestry management, managing urban green space, tourism management, robotics, and video and computer games' production. One of the main steps ...
Read More
Light Detection and Ranging (LiDAR) point cloud dataset and 3 dimensional (3-D) models have been extensively used for urban feature extraction, urban management, forestry management, managing urban green space, tourism management, robotics, and video and computer games' production. One of the main steps toward reaching accurate 3-D models is clustering and classification of LiDAR point clouds data. The main purpose of this research is to find out, particular machine learning techniques, which are promising for best learning and classification of LiDAR point cloud data in an urban area. Therefore, the performances of K-nearest neighbor (KNN), Decision Trees (D3), Artificial Neural Networks (ANN), Naive Bayes (NB), Support Vector Machine (SVM), and Markov Random Field (MRF) classifiers were evaluated on the LiDAR and aerial image dataset of Vaihingen, Germany, in the context of the "ISPRS Test Project on Urban Classification and 3D Building Reconstruction." In regard to the literature review, MRF model has not been used to classify LiDAR point cloud data in Iran. In this research, we utilized all the geometrical features, intensity values of LiDAR and aerial images as well as extracted eigenvalues based features to distinguish five urban object classes, including impervious surfaces, buildings, low vegetation, trees and cars. In order to compute eigenvalues using local point distribution, this paper introduces a new cubic structure, which has been not found in previous studies. The final results of 3D classification techniques in this research were 2D maps that evaluated by the benchmark ISPRS tests maps. The evaluation shows that the performance of MRF model with an overall accuracy of 88.08% and the kappa value of 0.83 is higher than other techniques to classify the employed LiDAR point clouds.
