SLIDE 1
1
2
Using sparse optical flow for multiple Kinect applications - - PowerPoint PPT Presentation
Using sparse optical flow for multiple Kinect applications - - PowerPoint PPT Presentation
Using sparse optical flow for multiple Kinect applications 27.6.2013 Stefan Guthe 1 Microsoft Kinect Recap and Impact Launched October 2010 Consumer-Grade RGB-D Sensor Over 3000 papers in the last 3 years related to the Kinect
SLIDE 2
SLIDE 3
3
Outline
Scenario: Gas capturing Related Work Our approach Algorithmic details Results Outlook
SLIDE 4
4
Scenario: Gas Capturing
Classical Approach: Schlieren Solved for 2d (Qualitative, Quantitative) 3D still relies on particles or Laser-Doppler Our scenario: Propane Gas travelling in air [Settles 2008]
SLIDE 5
5
Related Work
First approach to 3d Schlieren: Atcheson et al. 2008 Wavelet Noise Background Horn-Schunck Optical Flow to Detect Pixel deviations Reconstruction with Diffusion Tensor and Poisson Integration to 3d gradients in constrained Voxel volume [Atcheson 2008]
SLIDE 6
6
Our approach
Multiple Kinects
SLIDE 7
7
Our approach – the setup
Convergent multiple Kinect setup, common world space Projection walls reduce scene depth Index gradient by propane flowing with 4 bar in air
- ccluders with different aerodynamics
SLIDE 8
8
Our approach – gas detection in Kinect streams
First approach: Difference in depth image stream [Berger 2011] Second approach: IR stream + tailored optical flow
SLIDE 9
9
Algorithmic details
Mean spot intensity distribution of Kinect pattern resembles a aaussian Significant differences at nodge and in the tail regions
SLIDE 10
10
Algorithmic details
1) Find spots in reference image Bwmorph + brightness threshold 2) Fit mean spot kernel Least squares approach 3) For each frame and each spot Determine best fit of kernel within range from starting position Use least squares approach 4) Interpolate difference vectors to fill holes 5) Iteratively reconstruct volume based on difference vectors
SLIDE 11
11
Results
Outperformed State of the Art 2012 The PSNR of Kinect 1.0 is believed to be in the range 35dB - 20dB
SLIDE 12
12
Results – reconstructed gas flows
SLIDE 13
13
Conclusion and outlook
Employed multiple Kinects in convergent setup Reconstruction of Propane Flow in air around occluders Sparse Optical Flow based on Least-Squares Fit to spots Future work Compare Time-Of-Flight approaches to solution in Kinect 2.0 setups Mobile ad-hoc gas capturing (e.g. DARPA robotics challenge)
SLIDE 14
14
Fin.
Thank you for your attention
SLIDE 15
15
Multiple Kinect Studies
Project Sumary
! Capture scene motion data from
multiple unsynchronized RGB-D footage
! Heterogenous Sensor Setup ! Opaque and Transparent Object
Motion
! Summarization in 3D space
Published in VMV 2011, CDC4CV Schematic of Multiple Kinect Setup Hardware Shutter and Kinect
SLIDE 16
16
Multiple Kinect Studies – Heterogeneous Sensor Calibration
SLIDE 17
17
Tomographic Gas reconstruction
Project Sumary
! Record heated air flow atop camping
stove
! Rely on Background-oriented Schlieren ! Masking on optical flow ! Reconstruction based on radial basis
functions and diffusion tensors
! Poisson-Integration leads to Refractive
Index field Published in Springer Lecture Series
SLIDE 18
18
Tomographic Gas reconstruction
SLIDE 19
19
Multiple Kinect Studies – NiTe vs. Ours 1/2
SLIDE 20
20