The method of acquiring and processing 3D data from drones

Abstract

Rapid development of machine learning techniques opens new application fields for Unmanned Aerial Vehicles technology, which include detection and classification of objects. It is possible to detect buildings, vehicles or various objects present near pipelines and industrial buildings. In some cases, such as monitoring of the critical infrastructure, accuracy of the detection is crucial. 2D data classification enables detecting an object and determining its basic parameters. 3D data, that can be obtained from drones, supplement 2D data, and can significantly increase the accuracy of detection and classification of objects. It also bares additional information and can simplify determination of dimensions of already classified objects. Furthermore, some objects, difficult for classification using 2D images, can be easily classified with 3D data. Such objects are for example: excavations in the ground, objects partially overshadowed by trees or fully covered by dried leaves. 3D data collected by drones is typically obtained with SfM (Structure from Motion) and Lidar (Light Detection and Ranging) methods. SfM provides three-dimensional data from the photos that have been collected for 2D analysis. The advantage of this method is high quality texture. The main problem is that this method is not useful for night flights due to lack of feature points on images. Lidar is a laser measurement method using data on the time of flight of a laser beam reflected from an obstacle (object). It allows to obtain 3D data in all light conditions. However, collected data does not have color information. The combination of both methods will provide dense and accurate point clouds with texture, which can be consequently used for detection and classification of objects. In this paper a pipeline for acquisition, merging and processing of 3D data gathered by drones is presented. The first step is to obtain assembled point clouds from Lidar in one coordinate system using GPS data. Then Lidar point cloud is integrated with SfM point clouds. 3D data generated this way also includes coordinates of camera in the moments when SfM photos were collected. The full 3D model of monitored area containing GPS coordinates and positions of camera may be used to simplify configuration of a supplementary flight in order to measure places where no measurement data was obtained or the density of point cloud was too low. Having a point cloud of the reconstructed object prepared in such way, it is possible to compare point clouds, features extracted from point clouds and geometry of already classified objects over time.