紀要論文

GPUを利用した高速・高効率な３Dスキャナ

High-performance 3D Scanner using GPU

jpn

18810667

https://doi.org/10.15002/00010536

法政大学大学院情報科学研究科

9

143

148

2014-03

Version of Record

This paper proposes a high-performance real time 3D scanning method using GPU by extending Octree KinectFusion. In the Octree KinectFusion, all data are stored in video memory and a surface generated by a ray marching method can be displayed in real time. To make the data area larger, this paper proposes a method to use not only video memory but main memory. A way to use both main and video memories is similar to a swap system in operation system. When values on video memory get unnecessary for processing, they will be moved to main memory like swap-out and when values on main memory get necessary for processing, they will be back to video memory like swap-in. It costs so high for a program on GPU to exchange values on main memory that the proposed method inserts margins between camera screen and active video memory areas for avoiding excessive swap-in/out operations. To generate a surface faster, this paper uses a marching cube method. A ray marching method used in the previous study generated a depth map. It ran faster than a marching cube method but needed to run for every frame. On the other hand, a marching cube method generates polygons which neither changes often nor needs to run for every frame. Experimental results showed that the proposed method could capture bigger volumes of scenes and processed faster than the previous study and output high quality 3D model.

Article

AA12222297