--------------------------------------------------------------------------------------------------------------------

                          Visual Processing For Cross-Country Autonomous Navigation 
                     

                                      Roberto Manduchi
                                     UC Santa Cruz

--------------------------------------------------------------------------------------------------------------------

Terrestrial autonomous vehicles rely heavily on visual sensors for
environment perception ("world modeling"). For urban or arid terrain, the
main task of world modeling is to detect obstacles, by means of range
sensors such as stereo cameras or laser rangefinders (lidar). However, in
the case of cross-country vegetated terrain, the notion of "obstacle"
needs to be revisited. For example, small bushes or tufts of tall grass may
look like obstacles based on pure geometry description, yet they are
traversable by a suitably sized vehicle. In this talk, I will present visual
processing techniques for "terrain perception", which allow us to move from the
classical obstacle avoidance paradigm to a more efficient strategy of
obstacle negotiation. In particular, I will describe the use of visual
features such as color, texture, range texture and multispectral analysis,
to determine the composition of the terrain cover in outdoor vegetated
environments. I will then briefly discuss! how the terrain perception
module integrates into an end-to-end system for path planning and velocity
control.

This work was carried out at at the NASA Jet Propulsion Laboratory and
at UC Santa Cruz, as part of the XUV DEMO III effort and of the DARPA
Mobile Autonomous Robots Software program.