90 likes | 418 Views
The Stixel World – A Compact Medium Level Representation of the 3D-World. Hernan Badino , Uwe Franke , and David Pfeiffer Daimer AG DAGM 2009 Presenter: Jonghee Park GIST CV-Lab. Introduction. Stereo vision play an essential role for scene understanding in cars of the near future
E N D
The Stixel World – A Compact Medium Level Representation of the 3D-World Hernan Badino, UweFranke, and David Pfeiffer Daimer AG DAGM 2009 Presenter: Jonghee Park GIST CV-Lab.
Introduction • Stereo vision play an essential role for scene understanding in cars of the near future • SGM • Three out of ten most powerful stereo is SGM variants in Middlebury dataset • Implement using FPGA • Extract and track every object • Use multiple object detectors • Repetitively evaluate images • Incomplete detection
Introduction • Requirements of automotive environment perception and modeling • Compact • Reduction of the data volume • Complete • Information of interest is preserved • Stable • Small changes of the underlying data must not cause rapid changes within the representation • Robust • Outliers must have minimal
Free space • Occupancy Grid • Two dimensional grid which models occupancy evidence of the environments • Only 3D measurements lying above the road are registered • Free space computation [1] • The space found in front of the occupied cell is considered free space • Bottom of the obstacles is obtained from the free space [1] Hern´an Badino, Uwe Franke, Rudolf Mester, Free Space Computation Using Stochastic Occupancy Grids and Dynamic Programming, ICCV workshop 2007
Height Segmentation • The height of the obstacles is obtained by finding the optimal segmentation between foreground and background • Membership value • Depth difference between disparity • : bottom point set from free space analysis • : X,Z world position corresponding to • : set to 2 meter
Height Segmentation • Cost image from membership value • Accumulate from top to bottom • : the row position of lies on the road
Height Segmentation • Solve using dynamic programming • : 5m : possible same object depth • : 8 Smoothness term
Result • 25 ms on Intel Quad Core 3.00 GHz