Terrain perception for DEMO III

Abstract

The Demo III program has as its primary focus the development

Learning to Drive Among Obstacles

“…Obstacles are given a clearance which is wider than the clearance afforded by extreme routes. When clearance is not available, the algorithm plans at slower speeds [9]. Sandstorm and H1ghlander, robots developed for desert racing, have driven extreme routes at speeds up to 15 meters per second by planning in a series of grids along the original path and smoothing the result [10].…”

Fast and reliable obstacle detection and segmentation for cross-country navigation

“…These are compared against a prior obstacle detection technique [Bellutta00] that used slope measurements along image columns and 2-D area measures for obstacle detection. As mentioned earlier, this prior technique is not expected to work well on terrain that contains obstacles that slope along the image plane, rather than vertically downwards along image columns.…”

“…Our new rule-based 3-D shape reasoning and classification is expected to outperform prior 2-D based obstacle reasoning methods [Bellutta00] that used 2-D area information (not 3-D geometrical measures) to reject smallsized false obstacles. In many cases, an object that occupies a small number of pixels in the 2-D image does not imply the presence of a false obstacle, or conversely a camera could occupy a large number of pixels, and true obstacles far away from the camera could occupy a significantly small area.…”

“…Their technique (adopted by the DEMO III eXperimental Unmanned Vehicle (XUV) [Bellutta00]), measures slope and height of surface patches (where "slope" is measured by the angle formed by the surface normal and the vertical axis.) Figure 2 shows an example of a 1-D range profile, where slant (θ) and height (H) are shown for two different surface patches.…”

“…Figure 5 shows the process of segmentation/labeling of obstacles that occurs implicitly in our 3-D obstacle detector. Figures 6 and 7 show examples of a synthetic and real example where disparate objects that touch in 2-D space, and therefore assigned to one obstacle by a 2-D blob coloring procedure [Bellutta00], are labelled with different colors using our OS algorithm.…”