Kinect Localization and Mapping (Video 2) - iC2020

iC2020 - Fourth Year University of Waterloo Design Project
February 10th, 2011

Sean Anderson, Kirk MacTavish, Daryl Tiong, Aditya Sharma

Our goal is to use PrimeSense technology in order to create a globally consistent dense 3D colour map.

Future progress will be uploaded as it is completed!

Current accomplishments:
- Optical Flow using Shi Tomasi Corners
- Visual Odometry using Shi Tomasi and GPU SURF
. . . . Features undergo RANSAC to find inliers (in green)
. . . . Least Squares is used across all inliers to solve for rotation and translation
- Loop closure detection using a dynamic feature library
- Global Network Optimization for loop closure

Future steps:
- Thinning of global map using voxels
- Detection of "safe" terrain for driving

Underlying Technology:
- Basic ROS Utilities
- OpenCV
- GPUSURF
- TORO (openslam.org)

Category:

Science & Technology

Tags:

• Kinect
• SLAM
• robot
• robot mapping
• visual odometry
• SURF
• mapping
• 3D
• ROS
• OpenCV

[roidroid]

It's so finely detailed!

Now that you have loop closure, it's truly epic. Great work man!

Question: Is this all calculated in realtime?

[shyxguy]

@roidroid it is realtime =) our goal is to be able to navigate a robot around the room safely

[check86]

Awesome! Question: How are you determining the spatial position of the Kinect and how accurate is it?

[shyxguy]

@check86 its done through a combination of visual odometry and loop closure. There's a better overview of all the techniques in in the video description, you can google any of the parts and find lots of information on them. In terms of accuracy we haven't measured it yet

[SnakePlisskenPMW]

whats the final product... in what forms can you exsport...

[shyxguy]

@SnakePlisskenPMW our current goal is robotics oriented (motion estimation and object avoidance), but it would be interesting to use the models for other applications. Currently the format is just raw points (millions of x,y,z,r,g,b), I've never used UNITY or Blender, do you know if they include algorithms that would create a mesh from the point cloud?

[FatLingon]

This is awesome. But one question, is there unique technology inside the Kinnect, or would this be possible with any HD webcam or even a video-clip in retrospect?

[Madsy9]

@shyxguy You could probably convert all the points to polygons with delaunay triangulation, and generate texture maps and texture coordinates out of the vertex/point positions and colours. Not sure if you would get acceptable results without human intervention in the process, though.

[TheNordbjerg]

This is pure awesomeness :)

[artec1984]

Is this available for download?

[djkaspy]

Siiick