A 3D reconstruction from real-time stereoscopic images using

A 3D reconstruction from real-time
stereoscopic images using GPU
GOMEZ-BALDERAS, Jose-Ernesto, GIPSA-lab,
[email protected]
HOUZET, Dominique, GIPSA-lab,
[email protected]
Abstract
We propose 3D reconstruction method that uses a Graphics Processors Unit (GPU) and a disparity map from block matching algorithm (BM).
Context
Recent trends show that there is a high demand of 3D imaging in:
•Media and entertainment,
•Defense and Security,
•Architecture and Engineering.
3D technology is being implemented in various objects as it provides a
more realistic view than 2D:
•Machine vision
•Image segmentation for object recognition
•Defense and security via its usage in simulation
•Facial identification and target detection
Strategy
Our algorithm uses two stereoscopic video sequences like inputs and then it processes the two stereoscopic images using a GPU and then we can
visualize a 3D reconstruction in real-time.
Methods
a) Capture Images and Color to Grey Conversion :color
stereo images pair on RGB space are converted into
grey space images pair.
b) Sobel Filter and Rectified Stereo Images: we have
reduced our search in 1D using the epipolar geometry
(el=er) of the two images.
c) Stereoscopic Block Matching Algorithm in GPU On
CUDA each thread performs the computation to obtain
the disparity map dmB.
d) Reproject disparity map to 3D: a point in 2D can be
reprojected into 3D dimensions given their coordinates and
the camera intrinsic matrix.
e) 3D reconstruction visualization of point clouds: we
calculate the 3D coordinates of each point using the
disparity map dmB and we use a cloud point structure,
PC(i)={Xi, Yi, Zi, Ri, Gi, Bi} to visualize in real-time.
Results
The computer we used in experiments is equipped with
an Intel Core i7 3.07GHz, 5GB memory.
Conclusions
Experimental results show a speedup factor (4x faster) of our system
in contrast to CPU system. In addition, the achieved speedup shows
the importance of parallel algorithms and computing architectures in
GPGPU. With real-time performance, our system is suitable for
practical applications.
CPU
Intel Core i7
3.07GHz
FPS
IUJW_Left
IUJW_Right
Jamie2_L
Jamie2_R
129
GPU
NVIDIA
GeForce
GTX 285
FPS
291
98
290
Grenoble Images Parole Signal Automatique
UMR CNRS 5216 – Grenoble Campus
38400 Saint Martin d’Hères - FRANCE
2x
GPU
NVIDIA
Quadro
4000
FPS
413
3x
409
Speed
Factor
Speed
Factor
3x
4x