Quality-driven autonomous 3D reconstruction of large-scale scenes

点击翻译按钮获取中文摘要

Monitors and TVs have been the dominating display devices in the past 30 years. They are only capable of showing 2D contents, which can be easily acquired using today's digital cameras. As virtual reality (VR) devices gain practicality and popularity, allowing users to view 3D contents from selected viewpoints and directions, an important question is, therefore, how to efficiently and economically capture objects and scenes in 3D.The technology for acquiring 3D shapes of small objects is already mature. Users can scan a given object from different sides using a range scanner and then register together the 3D point clouds to generate a surface model. On this front, my collaborators and I have developed a state-of-the-art algorithm that automates the scanning process through positioning a scanner at strategically selected locations using a robotic arm. However, how to acquire high-quality models for large-scale outdoor scenes in an autonomous manner is still an open problem.3D reconstruction for large sites has vast applications in areas such as urban planning, geological and archaeological survey, virtual tourism, and military simulation. It is an active topic in many research communities, including computer vision, graphics, robotics, civil engineering, and remote sensing. In the past, people often resorted to manned airborne LiDAR systems, which come with intimidating cost. Thanks to the rapid development of unmanned aerial vehicles (UAVs), it is now possible to scan large sites in a much more economical manner.The proposed research aims at developing novel, efficient, and autonomous techniques for large-scale 3D scene reconstruction. To achieve this goal, in the next five years my students and I will investigate: 1) how to perform quality-guided autonomous 3D reconstruction for large outdoor scenes with LiDAR-equipped UAVs; 2) how to achieve the same objective using low-cost off-the-shelf UAVs with only image capture capability; 3) the benefits of letting both types of UAVs work collaboratively and fusing together image and range data; and 4) the feasibility of segmenting moving objects from point clouds and reconstructing dynamic 4D spatial-temporal models for these moving objects.As VR devices get widely adopted, the demands for large-scale scene models are expected to increase dramatically. Hence, the techniques to be developed in this research will make broad impacts in aforementioned areas and benefit different research communities. They are likely to advance the Information Systems and Technology field through making high-quality 3D models of our cities and environments easily obtainable. A number of research questions need to be addressed, which will help to train HQPs to take up highly demanded positions in Canada's high tech industries.