Automatic Perception and Target Detection in LiDAR Data

LiDAR is a remote sensing technology which uses a set of 3D geo-referenced points in order to describe a scene. Aerial LiDAR is often collected using UAVs or airplanes which can passively collect data over a short period of time, often over several miles. This can result in millions of points used to describe a scene. LiDAR data is often used for surveillance and military applications and because of the large amount of data and varying resolutions it can be difficult for analysts to recognize and identify mission critical targets within the scene. The goal of this project is to develop a technique for the automatic segmentation and classification of distinct objects within the scene to aid analysts in scene understanding. We focus our method on five distinct classes that we wish to identify; ground, vegetation, buildings, vehicles and fences or barriers. The first step is to use a RANSAC-based ground estimation in order to estimate the digital terrain model (DTM) of the scene. Next, 3D octree segmentation is performed in order to distinguish between individual objects within the data. A novel volume component analysis (VCA) method is used to extract distinct geometric signatures from each individual object and these features are used as the input to several support vector machines (SVM) in cascade of classifiers configuration. The cascade of classifiers separates the objects into the four remaining classes. Our method was tested on an aerial urban LiDAR scene from Vancouver, Canada with a resolution of 15.6 pts/m^2 and was found to have an overall accuracy of 93.6%.https://ecommons.udayton.edu/stander_posters/1650/thumbnail.jp