Intelligent Terrain Extraction From Airborne Lidar And Dense Stereo Matching Point Clouds

Currently,3D point cloud data acquisition technology has been an important technological innovation in remote sensing and mapping science. Airborne LiDAR and dense stereo matching point clouds are capable of directly acquiring high-accuracy 3D coordinates of discrete points on the earth’s surface. The extraction of Digital Elevation Models (DEM) from 3D point clouds have been an attractive research topic. As both ground and non-ground objects (e.g. buildings, vegetation, and vehicles) are included in the point clouds, the first step towards generating a DEM is classifying the point cloud into ground and non-ground points. This task is referred to as filtering. With the variations in scene complexity, point clouds filtering can be a difficult task, and its quality control procedure may consume approximately 60% to 80% of the total processing time. Moreover, filtering is also challenging for real-time applications or large-volume data processing.This paper aims to propose intelligent and fast algorithms for airbore LiDAR and dense matching point clouds filtering. The main research contents include:1)Explore quick and efficient noise reduction method for 3D point clouds. We propose a noise reduction method based on discrete roughness estimation. This method uses TIN to estimate the local roughness of all the points, and extcats noise points based on the roughness. This noise reduction method is fast and can take into account the smoothness of the point clouds.2) Scene knowledge guided 3D point cloud data filtering algorithm. The algorithm can automatically identify the scene knowledge and topographical features, based on which the filtering parameters are adaptively adjusted.3)Fast and high-quality filtering algorithm for 3D point clouds. This paper presents a fast and intelligent filtering algorithm. Inspired by the Semi-Global Matching (SGM) algorithm, we model the filtering task as a labeling problem in which an optimal classification surface close to the bare ground is computed by minimizing a novel energy function. The classification surface is then used to classify the points into either ground or non-ground. Energy function is optimized based on a semi-global search using Dynamic Programming (DP) from multiple 1D directions, it can be implemented by parallel or GPU computing.4)Filtering algorithm for dense matching point clouds. We present a dense matching point clouds filtering algorithm based on scan line segmentation. In order to improve the discrimination of the ground points and non-ground points, the algorithm uses the cross line element to segment the point clouds into different segments. MRF is then used for the classification.