Robotic-Vehicle Perception Control for Detecting Obstacles

Unified Approach to Control of Motions of Mobile Robots

Obstacle-detection systems are designed to make the most of limited data-processing resources.

A concept of perception control is guiding the continuing development of obstacle-detection systems for crosscountry navigation of robotic vehicles that are equipped with stereoscopic machine-vision systems. Perception control consists of optimally tuning sensor or processing parameters to increase efficiency of perception under design constraints and design requirements while adjusting to the environment. This particular concept of perception control is oriented toward the need to maximize vehicular safety at a given speed or, conversely, to determine the maximum speed for a given level of safety.

An obstacle-detection system according to this concept uses computing resources efficiently, without resorting to "brute-force" obstacle-detection techniques that often involve more computation than is necessary. Such a system is designed to implement a focus-ofattention approach, in which data are processed from subwindows of the stereoscopic video images of the path ahead, instead of from the entire images, to reduce the computational cost of perception. The image data are processed in the following main steps:

Pyramids (in a symbolic sense) of versions of the stereoscopic images that are band-pass-filtered in a succession of spatial-frequency bands are constructed from the stereoscopic pairs of images.
Cross-correlations are computed on any single level of an image pyramid to estimate stereoscopic disparity at every pixel of the pair of images.
The range (that is, the distance from the video cameras on the robotic vehicle) is computed from the disparity at every pixel
An obstacle-detection algorithm is applied to the resulting range image.

The obstacle-detection algorithm assumes that an obstacle consists of a nearly vertical step displacement on an otherwise nearly flat ground plane. The algorithm does this by using pairs of pixels in the same column of the range image (see figure). If the difference in height between the two pixels in any such pair exceeds a prescribed step height, then an obstacle is deemed to exist at the affected location.

The obstacle-detection system manages the computing resources by adjusting variables at three levels: image resolution, subwindows of attention (steps 2, 3, and 4 can be performed in subwindows), and detection threshold (which can be varied over the scene). The system takes into account its current lookahead requirements and determines the part of the path ahead that must be examined for obstacles at each processing step. The system assumes that the vehicle must stop before colliding with an obstacle, and allowance must be made for image-acquisition and processing time, actuation delay before brakes can be applied, and the braking time to reduce the speed from its current value to zero.

A velocity controller could be designed to operate in conjunction with an obstacle-detection system of this type. The velocity controller could be designed to try to reach the maximum allowable speed, provided that it always checked for the distance to the end of the last processed path segment and applied brakes if necessary (that is, if data on the next path segment did not come in time).

Point of Contact:
Larry Matthies
Mail Stop 107-102
Jet Propulsion Laboratory
4800 Oak Grove Drive
Pasadena, CA 91109
818-354-3722
Larry.H.Matthies@telerobotics.jpl.nasa.gov

Telerobotics Program page

Please email the site webmaster with any comments, criticisms or corrections for this page.
Maintained by: Dave Lavery
Last updated: May 10, 1996