PixelLaser: Range from texture

PixelLaser: Range from texture
ISVC '10
11/30/2010
Las Vegas, NV
Max Korbel ’13, Michael Leece ’11, Kenny Lei, Nicole Lesperance ’12, Steve Matsumoto ’12, and Zachary Dodds
Motivation
Pipeline
From their earliest days (e.g., Horswill’s Polly) robots have used
image segmentation to estimate which way to steer next, i.e., the
general traversability of the terrain ahead. This project pushes
segmentations one step further – to build range scans similar to
laser range finders (LRFs). Our approach seeks to make LRFs’
large body of mapping, localization, and navigation algorithms
available to a much wider audience through low-cost platforms.
These image-segmentation scans can then serve as the
basis for off-the-shelf spatial reasoning algorithms such as
localization and mapping.
Scans from Segments
Classification
Image descriptors
and their redundancy
Original Image
Segmentation
We use nearest-neighbors classification
on small image patches to determine
traversable from untraversable texture.
A comparison of the color and texture
filters, shown at left, has guided the
selection of image-patch descriptors.
Training: a training image and the patches indexed in
the Kd-tree. Blue (red) patches are (un)traversable.
The transformation from segmentation
to distance depends on the height,
angle, and internal geometry of the
camera. Rather than calibrate, we
empirically fit a function mapping from
Plot of range vs. row
image height to range-to-obstacle.
Mapping
Just RGB Statistics
RGB and Texture Filter
Examples of classified patches
Classification: the nearest neighbors of one patch and
the overall results of classifying a novel image.
Coreslam results from the “playpen”
Our Python port of
CoreSLAM yields
maps of a quality
the same as the
original authors’.
Localization
Our implementation of Monte Carlo Localization using imagesegmentation-based scans shows their power and promise.
Each image is segmented via a multi-resolution search for the
bottommost transition from traversable to untraversable texture.
We are investigating genetic-algorithm approaches to find the
patch-descriptor weights that best segment our images.
Currently the most expensive piece of the procedure is the
nearest-neighbor lookup within the large K-d tree of remembered
patches.
Elementary!
Application
Platform
This project uses a netbook with OpenCV
and Python atop an iRobot Create. The
robot is robust and flexible enough to be
our primary outreach platform, too. Note
that a LRF would cost many times more
than this entire platform!
Segmentation: we run
at several resolutions to
search for transitions in
terrain traversability.
Range scan: the
resulting range scan,
shown here as it would
look in a top-down view.
PixelLaser-based MCL
Acknowledgments
We gratefully acknowledge support by The Rose Hills
Foundation, Baker Foundation, the NSF projects REU
#0753306, CPATH #0939149, and funds from HMC.