Analyzing 3D point clouds, generated by various cameras, to support autonomous vehicle navigation

LENZANO, MARÍA GABRIELA

Conferencia; COUNT meeting; 2018

Autonomous vehicles require high accuracy maps enabling precise positioning on roads based on reference points on 3D models of the environment. A safe navigation through the environment requires a detailed knowledge of the 3D space surrounding, then, high-definition map data is an essential issue to provide information to localize objects, determine the location, speeds, etc. In recent years, photogrammetry and computer vision have sought to improve automated image matching where matched feature points in overlapping photographs enable the derivation of 3D coordinates as point clouds. Structure from Motion (SfM) is a 3D point cloud generation technique that achieves highly accurate solutions in different applications at a low cost. SfM photogrammetry can be performed with images acquired by consumer-grade digital cameras and is thus very flexible in its implementation. In this study, we investigate how cameras such as DSLR and fisheye lens (GoPRO) can be used for urban environments dense 3D point clouds, comparing the performance the solutions, and evaluating tools (VisualSfM and PIX4D). Our goal is to provide a detailed examination of the ability of Structure-from-Motion photogrammetry to represent an accurate 3D surface on urban environment, and to reliably detect small details. We expect to obtain the 3D data with decimeter accuracy which is adequate to achieve a safe navigation. The test was carried out at the campus of The Ohio State University, and the cameras mounted in a mobile acquired more than 2,000 images. Our results show that SfM fully automated data processing creates both opportunities and limitations, particularly, using data captured by a mobile in movement in urban environments.