Your SlideShare is downloading. ×
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Lugano2
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×
Saving this for later? Get the SlideShare app to save on your phone or tablet. Read anywhere, anytime – even offline.
Text the download link to your phone
Standard text messaging rates apply

Lugano2

424

Published on

0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total Views
424
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
6
Comments
0
Likes
0
Embeds 0
No embeds

Report content
Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
No notes for slide

Transcript

  • 1. Calibration-less Anthropometric Scanner Using GPUs Mario Gazziro, Ph.D.P. Scotton, H. Bittencourt, A. Osti, F. ColaUniversity of Sao Paulo - BRAZIL
  • 2. Anthropometric 3D Scanners ApplicationsBody Mass Index determinationThe scanner can determine accurately the volume of the patient and(with his weight)could obtain the density d.Using the density value obtained, wecan evaluate the percentage of bodyfat (%BF) using the Siris equation:%BF = (4.95 / d - 4.50) * 100d: body density (g/cm3)
  • 3. Anthropometric 3D Scanners ApplicationsPostural Analysis AssessmentPostural analysis are used in standing postural measurements in acompletely automated way, replacing the symmetry grid by the 3Dscanner.
  • 4. Goals Developing an Anthropometric 3D Scanner with the following features:● Fast operation, allowing to generate profiles (BMI and Postural Analisys) from large populations in a short time (using GPU acceleration)● Allow its installation, maintenance and usage by people not specialized in computer graphics (regular technicians from hospitals, clinics and gyms)
  • 5. Calibration-less challengeTo avoid the usage of calibration patterns in amulti-sensor approach, we try two algorithms toautomatic registering the cloud points:– EM-ICP algorithm– Softassign algorithmBoth algorithms could not register point cloud of thehuman body from different points of view, even withreasonable overlap (about 70 clouds/270o).
  • 6. Approach: Use only ONE sensor andSLAM techniques to scan the body * Simultaneous Localization And Mapping
  • 7. Prototype developed using SLAM
  • 8. SLAM implementation: KinFu● OpenSource implementation of Microsoft tm tm Kinect Fusion API (filters for SLAM)● Supported by PCL – pointclouds.org
  • 9. GPU used to accelerate KinFu● Model: GTX 680● Cuda Cores: 1536● Memory: 2 GB DDR5● PCI Express: 3.0● Architecture: Kepler GK110
  • 10. NVidia Kepler GK110 Architecture
  • 11. TSDF CloudsTSDF stands for Truncated Surface Distance Function and was first introduced byBrian Curless and Marc Levoy in A Volumetric Method for Building Complex Models fromRange Image paper in the Proceedings of SIGGRAPH 1996. “What is the difference between a TSDF cloud and a normal point cloud?”Well, a TSDF cloud is a point cloud. However, the TSDF cloud makes use of how the data isstored within GPU at KinFu runtime.The cube is subdivided into a set of Voxels. These voxels are equal in size. The default sizein meters for the cube is 3 meters per axis. The default voxel size is 512 per axis. Both thenumber of voxels and the size in meters give the amount of detail of our model. * pictures and text from from pointclouds.org
  • 12. TSDF Volume Grid in GPUA representation of the TSDF Volume grid in the GPU. Each element in the grid represents avoxel, and the value inside it represents the TSDF value. The TSDF value is the distance tothe nearest isosurface. The TSDF has a positive value whenever we are “in front” of thesurface, whereas it has a negative value when inside the isosurface. The X,Y,Z coordinatesfor each of the extracted points correspond to the voxel indices with respect to the worldmodel.
  • 13. Anthropometric Scanner at work
  • 14. Surface Reconstruction: PoissonParameters:Octree Depth factor 10and Solver Divide factor 8Grid based method,very good automaticresults, but withhuge memoryutilization32 GB RAM in ourdesktop
  • 15. Body Mass Index determination
  • 16. Postural Analysis* Grant R. Tomkinson *, Linda G. Shaw, Quantification of the postural and technical errors in asymptomatic adultsusing direct 3D whole body scan measurements of standing posture. Gait & Posture, Elsevier, In Press.
  • 17. Postural AnalysisJacques Junior, J. et al. Skeleton-based human segmentation in still images. ICIP, 2012.Their algorithm using our skeleton and RGB image to detect posture (red markers)
  • 18. ConclusionsIn december next we will start aprotocol with 160 subjects, to evaluateBMI and postural analysis.The postural analysis will be validatedagainst a specialist visual inspection.The BMI tests will be validated withrespect to the pletsmography(BOD-POD show in the figure).Pepper et al. already validate the BMImethod with respect to the hidrostaticweighing and DEXA (Dual X-Ray)methods, with an error of 1,5%, usinga 3D scanner from CyberWare andtesting 70 subjects.We will try to improve this results,onceour system is faster and have lessproblems due to breath and movementof the subjects. Pepper et al., Evaluation of a rotary laser body scanner for body volume and fat assessment. J. Test and Eval, NIH, 2010.
  • 19. So long and thank you for all the fish!

×