Next: Robot Navigation in Up: Active Perception Previous: A Flexible Behavioral

Visual Collision Detection for a Mobile Robot

Authors: [tex2html_wrap4186]A. Lin, M.D. Levine

Investigator username: levine

Category: perception

Subcategory: active perception

Within the framework of active vision, we are concerned with the task of allowing an autonomous mobile robot to move freely about without collision in its environment, guided by a video camera. Our approach is motivated by neurophysiological findings that, in primates, neurons in the medial superior temporal (MST) area of visual cortex respond preferentially to global translational, rotational, and expansive/contractive motion. The sensitivity to each of these varies from cell to cell, suggesting that MST neurons are specialized for detecting different types of motion. The cells integrate information over much of the visual field, and hence may play a significant role in the detection and analysis of wide-field motion; such motion is induced predominantly by eye and head movement, but also by large or proximate moving objects.

Ultimately, we wish to design three detector units, one for each of translational, expansive/contractive, and rotational motion. The units comprise two stages, the first being a bank of local velocity detectors. In biological terms, the functionality of this bank corresponds to that provided by the middle temporal (MT) area of visual cortex. However, due to the motion aperture problem, only the normal components of velocity are measurable; we recover an estimate of the true velocity using the Hough transform technique.

The outputs of this first stage velocity computation are further processed to obtain wide-field motion information. For our specific problem of detecting looming objects, we concentrate on expansion detection. This involves both localization and rate-of-expansion estimation, which can be done simultaneously through an iterative optimization scheme.


baron@cim.mcgill.ca