Back
to indexCoombs, Herman, Hong, Nashman
Summary
Good, intuitive algorithm for obstacle detection, but highly computationally expensive and extremely sensitive to noise and vibration between sequences of images. The mobile robot platform used worked at 30 cm/s, or .671mph, and the authors suggested this was too fast since the robot needed to take many samples as it spun around in order to increase accuracy. The concept is simple: use the relative image "flow" to detect divergences in the incoming image stream over time, and detect incoming obstacles using large central flow divergence around the obstacles to determine their precise location. Active mechanical gaze stabilization is required to effectively match pairs of images over time.
The interesting parts of the paper included the utilization of a "forgetting factor", which allowed the algorithm to adjust how relevant the previous image is to the current image, so when the robot is turning, and much of the scene is changing, it can decrease the importance of the previous image dynamically. Likewise, a robot moving forward decreases the forgetting factor, so the images become more important as time continues.
The authors also observed hardware constraints when adjusting steering, since there was a latency on the servo control, moving cameras, and the possibility of overshooting a target if the rotational velocity is too high. Gaze stabilization is imperfect, and can contaminate a series of images if a certain, unlikely forgotten frame contains significant vibration or jitter.
Methods
Divergence of image flow is the sum of the image flow derivatives in two perpendicular directions. It provides a robust metric for scene structure over time, and is inversely related to time-to-contact, which is related to the distance to the object. Single 2D images are used, taken from 2 different cameras: a 40degree FOV looking toward the ground, overlapping with a 115degree FOV camera looking ahead. A Real Time Control System (RCS) decomposes goals both spatially and temporally, consisting of parallel processing modules to accomplish sensory processing, world modeling, and behavior generation. Normal flow is computed first, from a gaussian filtered image, by calculating the temporal derivative. The maximum flow value in the receptive field indicates the distance to the nearest object in that field. This is a highly computational task, requiring multiple iterations over the same data to derive the desired calculations. The time to contact is then calculated, from the derivatives in the spatial translations of objects between temporal sequences. Recursive Least Squares is applied to update the predicated time to contact, and out pops the obstacle distances.
Keywords
Image flow, divergence, servo control, 2D time sequenced images
Rating
7
Bibtex Entry
@article{ debra94,
author = "P. M. E. De Bra and R. D. J. Post",
title = "Information retrieval in the World-Wide Web: Making client-based searching feasible",
journal = "Computer Networks and ISDN Systems",
volume = "27",
number = "2",
pages = "183--192",
year = "1994",
url = "citeseer.nj.nec.com/99604.html"
}