Back
to indexSteven Goldberg, Mark Maimone, and Larry Mattheis
Summary
Good paper detailing the stereo vision algorithms for the Mars Rover, which is due to land 2004 (right during the DARPA Grandchallenge Race in March). Many timing and baseline studies were performed to analyze and optimize the stereo vision algorithm performance. Various hardware components were measured, with the conclusion being the radiation hardened, slower CPU units were less desirable to their commerical counterparts, and the radiation impact is negligable for these systems. Finally the GESTALT Navigation system is explored, though the conclusions are rather intuitive, see below.
Methods
The stereo vision algorithm calculates the distance to a target using a pair of calibrated lenses and 1) downsample the image (depth resolution les precise by factor of 2) 2) rectification 3) compute Laplacian to remove pixel intensity bias 4) compute disparity range 5) apply checking algorithm 6) reduce invalid results 7) compute disparity on model. Optimizations were chiefly by gaussian and decimation procedures, pre-processing each row, optimize within an inner loop, and finally compute the sub-pixel disparities. The GESTALT Navigation system uses waypoints as goals, which calculates the difference from the current GPS location to the waypoint, then looks for obstacles, then chooses a safe direction (based on a goodness and certainty value). Its pretty intuitive, and unfortunately has very little value added since its so darn basic. Good optimization algorithms.
Keywords
stereo vision, vision optimization, Laplacian pixel intensity bias, GESTALT Navigation
Rating
6
Bibtex Entry
@article{ goldberg02stereo,
author = "Steven Goldberg and Mark Maimone and Larry Mattheis",
title = "Stereo Vision and Rover Navigation Software for Planetary Exploration",
journal = "IEEE Aerospace Conference Proceedings",
month = "March",
year = "2002",
url = ""
}