20-03-2012, 11:53 AM
Multi-cue Visual Obstacle Detection for Mobile Robots
[attachment=18605]
INTRODUCTION
OBSTACLE detection is one of the most fundamental
needs for an autonomous navigation system to work.
In order to avoid obstacles, the majority of approaches use
laser or sonic range sensor devices. While sonic sensors are
imprecise, short-sighted and usually unreliable, lidar devices
are expensive. Nowadays, autonomous robots are usually provided
with stereo camera pairs in order to perform tasks such
as object recognition or visual SLAM. Thus, by using camera
pairs also for obstacle detection, the cost of a laser device can
be saved up. While vision systems can detect objects in their
whole visual field, common laser sensors only sweep a plane,
so those objects not intersecting the plane are missed. Twoaxis
sweeping lidars or integrated image and lidar sensors are
even more expensive.
GEOMETRIC-BASED DETECTION
The geometric-based obstacle detection algorithm is based
on the idea that the floor is approximately planar. Given
this assumption, the floor induces a planar homography between
the two camera images. A planar homography is a
projective geometry transformation that maps projected 3D
points between two image planes assuming that they rest on
a particular plane[1]. Pixels are mapped by premultiplying its
homogeneous coordinates by the homography matrix:
COLOR-BASED DETECTION
The color-based obstacle detection is inspired on the approach
seen in [5], pixels are classified based on their presence
in a histogram which is obtained after a training process.
The training consists in the selection of several image regions
of the ground and the computation of a three dimensional
histogram from those image regions. Region selection can be
done manually by a human operator or, if the environment
obstacles are textured (e.g. there are no untextured walls), by
the robot itself selecting floor regions using the geometricbased
classification previously detailed.
MULTIPLE CUE OBSTACLE DETECTION
An obstacle detection system based on only one of the
described methods would perform well under certain circumstances:
planar floor and textured obstacles in the geometric
approach; and disjoint color sets for obstacles and floor in the
color-based one.
CONCLUSIONS
In this paper we have presented two improvements to
previous single cue detectors and a new technique for merging
their outputs. The new single cue detectors, appearance and
geometry based, outperform their previous counterparts.
[attachment=18605]
INTRODUCTION
OBSTACLE detection is one of the most fundamental
needs for an autonomous navigation system to work.
In order to avoid obstacles, the majority of approaches use
laser or sonic range sensor devices. While sonic sensors are
imprecise, short-sighted and usually unreliable, lidar devices
are expensive. Nowadays, autonomous robots are usually provided
with stereo camera pairs in order to perform tasks such
as object recognition or visual SLAM. Thus, by using camera
pairs also for obstacle detection, the cost of a laser device can
be saved up. While vision systems can detect objects in their
whole visual field, common laser sensors only sweep a plane,
so those objects not intersecting the plane are missed. Twoaxis
sweeping lidars or integrated image and lidar sensors are
even more expensive.
GEOMETRIC-BASED DETECTION
The geometric-based obstacle detection algorithm is based
on the idea that the floor is approximately planar. Given
this assumption, the floor induces a planar homography between
the two camera images. A planar homography is a
projective geometry transformation that maps projected 3D
points between two image planes assuming that they rest on
a particular plane[1]. Pixels are mapped by premultiplying its
homogeneous coordinates by the homography matrix:
COLOR-BASED DETECTION
The color-based obstacle detection is inspired on the approach
seen in [5], pixels are classified based on their presence
in a histogram which is obtained after a training process.
The training consists in the selection of several image regions
of the ground and the computation of a three dimensional
histogram from those image regions. Region selection can be
done manually by a human operator or, if the environment
obstacles are textured (e.g. there are no untextured walls), by
the robot itself selecting floor regions using the geometricbased
classification previously detailed.
MULTIPLE CUE OBSTACLE DETECTION
An obstacle detection system based on only one of the
described methods would perform well under certain circumstances:
planar floor and textured obstacles in the geometric
approach; and disjoint color sets for obstacles and floor in the
color-based one.
CONCLUSIONS
In this paper we have presented two improvements to
previous single cue detectors and a new technique for merging
their outputs. The new single cue detectors, appearance and
geometry based, outperform their previous counterparts.