Upload
others
View
6
Download
0
Embed Size (px)
Citation preview
Applied Computer Vision for Robo4cs – SS 2014
Freespace, Ground Plane Es4ma4on and IPM
Final Project Presenta4on Bug Busters
Applied Computer Vision for Robo4cs – SS 2014
Agenda • Task Descrip4on • Our Approach – Dense Depth Image – Ground Plane Es4ma4on – Digital Eleva4on Map – Free Space Es4ma4on – Inverse Perspec4ve Mapping
• Possible Improvements
15/07/14 Bug Busters – Final Project Presenta4on 2
Applied Computer Vision for Robo4cs – SS 2014
Task Descrip4on • Find the Ground Plane
• Perform a Free Space Es4ma4on
• Project the Free Space into Bird’s Eye View (Inverse Perspec4ve Mapping)
15/07/14 Bug Busters – Final Project Proposal 3
Applied Computer Vision for Robo4cs – SS 2014
Dense Depth Image
• 1st approach: Disparity image from Assignm. 4
• Results were lacking informa4on in road plane
15/07/14 Bug Busters – Final Project Proposal 4
Applied Computer Vision for Robo4cs – SS 2014
Dense Depth Image
• Improvement by using ELAS algorithm
• Op4mized for smooth ground plane
15/07/14 Bug Busters – Final Project Proposal 5
Applied Computer Vision for Robo4cs – SS 2014
Ground Plane Es4ma4on • Only data from lower image half is used to reduce influence from sky ar4facts
• Point cloud and prior knowledge are fed into RANSAC framework
• Extrinsic calibra4on of camera posi4on to 1.65 m above ground level
15/07/14 Bug Busters – Final Project Proposal 6
Applied Computer Vision for Robo4cs – SS 2014
Ground Plane Es4ma4on • 4 plane parameters: 𝑎𝑥+𝑏𝑦+𝑐𝑧+𝑑=0
• Extended Normal vector: 𝑛 =(█■𝑎@𝑏@𝑐@𝑑 )
• Distance calcula4on: 𝐷= 𝑃 ↑𝑇 ∙ 𝑛
15/07/14 Bug Busters – Final Project Proposal 7
Applied Computer Vision for Robo4cs – SS 2014
Digital Eleva4on Map (DEM) • DEM is a 2.5-‐dimensional grid where intensi4es represent highest points in n-‐direc4on at (x,z)
• Transforma4on to polar coordinates:
𝑟= √𝑥²+𝑧²
𝜑=atan(𝑧/𝑥 )
• Region growing algorithm to create denser representa4on of eleva4on data
15/07/14 Bug Busters – Final Project Proposal 8
Applied Computer Vision for Robo4cs – SS 2014
Digital Eleva4on Map (DEM)
15/07/14 Bug Busters – Final Project Proposal 9
Applied Computer Vision for Robo4cs – SS 2014
Free Space Es4ma4on • Free Space is es4mated in dense DEM image
• Propagated from front to back un4l an obstacle is reached
• Missing data is stepped over (within limits)
• Contour finding: Only the contour with the largest area is kept, others are false posi4ves
15/07/14 Bug Busters – Final Project Proposal 10
Applied Computer Vision for Robo4cs – SS 2014
Free Space Es4ma4on
15/07/14 Bug Busters – Final Project Proposal 11
Applied Computer Vision for Robo4cs – SS 2014
Inverse Perspec4ve Mapping • Backtransforma4on of Free Space into Cartesian coordinates
𝑥=𝑟sin 𝜑
𝑧=𝑟cos𝜑
• Overlay with pixel intensity DEM image to create Bird’s Eye View of situa4on (IPM)
15/07/14 Bug Busters – Final Project Proposal 12
Applied Computer Vision for Robo4cs – SS 2014
Inverse Perspec4ve Mapping
15/07/14 Bug Busters – Final Project Proposal 13
Applied Computer Vision for Robo4cs – SS 2014
Backprojec4on • Projec4ve transforma4on of polar Free Space
• Δα = const. for every pixel row
• Δr increases with α
• 𝑟= tan(90°− 𝛼) ∗ℎ
15/07/14 Bug Busters – Final Project Proposal 14
∞
Δα
α
Δr
Image plane (KITTI image)
pixel rows
h
r Ground plane (DEM image)
Applied Computer Vision for Robo4cs – SS 2014
Backprojec4on
15/07/14 Bug Busters – Final Project Proposal 15
Applied Computer Vision for Robo4cs – SS 2014
Possible Improvements • Further tuning of parameters for becer Ground Plane Es4ma4on
• Incorporate models for sloping roads (B-‐Splines)
15/07/14 Bug Busters – Final Project Proposal 16
Applied Computer Vision for Robo4cs – SS 2014
Thank you for your acen4on!
Bug Busters