A Learning System Approach for Terrain Perception Using Eigenimages. A Vision Based Technique for Classifying Terrain Types


  1. Broten, G.S.
  2. Digney, B.L.
Corporate Authors
Defence R&D Canada - Suffield, Ralston ALTA (CAN)
Unmanned group vehicles (UGV) traversing open terrain require the capability of identifying non-geometric barriers or impediments to navigation such as soft soil, fine sand, mud, snow, compliant vegetation, washboard, and ruts. Given the ever changing nature of these terrain characteristics, for a UGV to be able to consistently navigate such barriers, it must have the ability to learn from and adapt to changes in these environments conditions. In order to identify these non-geometric barriers a UGV must have a system that is capable of perceiving terrain types and determining the corresponding terrain properties. This research investigated applying the eigenimage algorithm to simple visual images, in conjunction with learning algorithms, to categorize terrain conditions. The attraction of this approach is that the algorithms are dynamic in nature and thus possess the capability of learn from experience. The experiment conducted so far have shown that the investigated algorithms were not able to consistently determine terrain conditions. The results of this research has lead to the speculation that eigenimages may not be well suited for the extraction of features from natural images. While the eigenimage algorithm did not perform at an acceptable level it is still believed that a learning based approach to the trafficability problem is the appropriate avenue of research.

Il y a un résumé en français ici.

Report Number
DRDC-SUFFIELD-TR-2003-165 — Technical Report
Date of publication
01 Dec 2003
Number of Pages
Document Image stored on Optical Disk

Permanent link

Document 1 of 1

Date modified: