Terrain Classification from 3-D Point Clouds for Autonomous Mobility


People



Description

This project investigates extracting information from 3-D point clouds generated by mobility sensors (such as scanning ladars) for unmanned ground vehicles.  The basic approach involves 1) computing features in a local neighborhood centered at each data point,  2) classifying each point using the local features, and 3) grouping the points into extended structures by enforcing spatial continuity.  The  features are based on the local spatial distribution of the dataand the local classifiers use standard statistical classifiers such as SVMs.

The algorithms developed in this project have been demonstrated on input from several different sensors and they were integrated in a complete unmanned ground vehicle system. Typical classification tasks include:

Terrain classification

Typical terrain

Classification output (red = load bearing surface; green = vegetation; blue = linear structures)
Object map generation

Input 3-D point cloud

Ouput object map
Wire detection

Input 3-D point cloud

Detected wire (in blue)
Detection of organized structures

Scene and input 3-D point cloud

Detected structure  (in blue)

Many aspects of the 3-D terrain classification problem are addressed in other projects. In particular, the 3-D point clouds are rapidly updated (100K points/sec or more) and the density of the point distribution may vary drastically across the point cloud. For these reasons, new tools were developed for efficient computation on dynamically-update point clouds, and for estimating the characteristic scale of  3- D point clouds for local feature computation. In addition, tools and features specifically designed for  structured environments (e.g., urban areas) are also investigated.

References

Jean-Francois LalondeNicolas Vandapel, Daniel Huber, Martial Hebert,  Natural Terrain Classification using Three-Dimensional Ladar Data for Ground Robot Mobility Journal of Field Robotics, Volume 23, Issue 10, 2006. 

Nick Heckman, Jean-Francois LalondeNicolas Vandapel, Daniel Huber, Martial Hebert,  Potential Negative Obstacle Detection by Occlusion Labeling.  In the Proceedings of the IEEE/RSJ Int'l Conf. on Intelligent Robots and Systems (IROS) , 2007. 

J. Tuley, Nicolas Vandapel, Martial Hebert, Analysis and Removal of Artificats in 3-D LADAR DataIn Proceedings IEEE International Conference on Robotics and Automation (ICRA), 2005. 

Nicolas Vandapel and Martial Hebert, Finding Organized Structures in 3-D Ladar Data. In the Proceedings of the IEEE/RSJ Int'l Conf. on Intelligent Robots and Systems (IROS) , 2004. 

Nicolas Vandapel, Daniel Huber, A. Kapuria,  Martial Hebert, Natural terrain classification using 3-d ladar dataIn Proc. IEEE Int'l Conf. on Robotics and Automation (ICRA) , 2004.

Funding

This research is supported by:

Copyright notice