1 / 16

Visual Odometry David Nister, CVPR 2004

Visual Odometry David Nister, CVPR 2004. 2005. 1. 4 Computer Vision Lab. Young Ki Baik. Contents. Introduction Algorithm Experimental results Conclusion and opinion. Introduction. Visual Odometry Usage of visual information as a sensor

gala
Télécharger la présentation

Visual Odometry David Nister, CVPR 2004

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Visual OdometryDavid Nister, CVPR 2004 2005. 1. 4 Computer Vision Lab. Young Ki Baik

  2. Contents • Introduction • Algorithm • Experimental results • Conclusion and opinion.

  3. Introduction • Visual Odometry • Usage of visual information as a sensor • Realization of the real-time navigation system using 3D reconstruction algorithms (camera motion estimation algorithm) • Features for real-time • Parallel processing based PC (MMX) • Pentium III 1GHz • Fast algorithm • Preemptive RANSAC (ICCV2003) • Features for accuracy • Stereo camera • Calibrated framework

  4. Introduction • System overview 3D reconstruction Feature extraction Motion estimation Matching and tracking 5-point algorithm / P-RANSAC / Triangulation method / Bundle adjustment Harris corner detector Normalized correlation 3-point algorithm for 3D motion

  5. Algorithm • Feature extraction • Harris corner detector • No subpixel precision detection • Usage of down sampled data (16 bit) • Size of INT and FLOAT is 32 bit. • Low size of data can be expected more efficiency for parallel processing. 32 bit MMX register 16 bit 64bit

  6. Algorithm • Feature matching • Normalized correlation over an 11x11 window • 11x11 = 121 (for applying to 128 bit aligned memory) • Matching with converted 1 dimensional vector using Parallel processing (MMX) is faster than normal method. • Short search range (Video sequences have short base line) 7 121 Garbage space … Matching using MMX … …

  7. Algorithm • 3D reconstruction • 5-point algorithm • Only considering pose estimation. • Usage of 2D points. • Preemptive RANSAC (CVPR 2003) • Fast RANSAC • Triangulation method • Conventional triangulation method is used for 3D reconstruction. • Bundle adjustment • Using small number of parameters and iteration.

  8. R, T Algorithm • Motion estimation • 3-point algorithm • Only considering camera pose (rotation and translation) estimation. • Usage of 3D point. Generated points Triangle Selected points

  9. Algorithm • Merit of using the Stereo Vision • Known scale (baseline) • Less affection by uncertainty in depth

  10. 3D motion (3-P algo., P-RANSAC) Motion estimation (5-P algo., P-RANSAC) Triangulation Stereo camera Matching Algorithm • The Stereo Scheme Triangulation Stereo camera Matching Next frame R, T

  11. Algorithm • The Stereo Scheme 3D motion estimation Certain number of frames Optimization (LM) Coordinate system is transferred. Firewall For stopping propagation error

  12. Experimental results • System configuration • CPU : Pentium III 1GHz (MMX) • Stereo camera • (360*240*2) size / FOV : 50˚ / Baseline : 28 cm • Experiments • GPS : Location error test • INS : Direction error test • Environment • Loop • Meadow • Woods

  13. Experimental results • Processing time • Around 13Hz • Location error • Direction error

  14. Experimental results • Performance Red line : Visual odometry Blue line : DGPS

  15. Experimental results • Performance Red : Visual odometry Blue : DGPS

  16. Conclusion and Opinion. • Conclusion • Real-time navigation system is implemented. • Opinion • There is no refinement scheme for solving closing loop problem. • More fast result with Pentium-IV (SSE2) • There is room for improvement.

More Related