100 fps & Very Low Drift Visual Odometry - New College Data Set (source code available)

The video shows the results of estimating visual odometry on the New College data set (http://www.robots.ox.ac.uk/NewCollege.... Visual odometry is obtained by an optimized version of the algorithm of Badino, Yamamoto and Kanade [1] and [2] that minimizes the reprojection error of features tracked and integrated over multiple frames. Features are tracked by the KLT algorithm and integrated over time by means of a Kalman filter.

Main window: left view.
Top-right: optical flow vectors of the tracked features. The color encodes the length of the optical flow vector.
Middle-right: color encoded stereo disparities from green to red.
Bottom-right: estimated traveled path by dead-reckoning (no loop closures).

The very low drift was achieved by an appropriate parametrization of the algorithm.

The whole computation time using one core of on an Intel(R) Core(TM) i7-5930K CPU @ 3.50GHz is 9.7 ms, achieving a frame rate of more than 100 frames per second.

The source code is available on SourceForge.net: http://sourceforge.net/projects/qcv/

[1] Hernan Badino, Akihiro Yamamoto, and Takeo Kanade. Visual Odometry by Multi-frame Feature Integration. In International Workshop on Computer Vision for Autonomous Driving (CVAD 13) @ ICCV 2013, Sydney, Australia.


[2] Hernan Badino and Takeo Kanade. A Head-Wearable Short-Baseline Stereo System for the Simultaneous Estimation of Structure and Motion. In IAPR Conference on Machine Vision Applications (MVA), Nara, Japan, June 2011.

Hernan Badino
http://lelaps.de/