People Detection and Tracking from Small-Footprint Ground Robots Tyler Lyons

advertisement
People Detection and Tracking from
Small-Footprint Ground Robots
Tyler Lyons
Overview
•
•
•
•
Introduction
Implementation
Experiments
Questions
Introduction
• Video!
Implementation
• Data Preprocessing
• Candidate Legs Selection
• Legs Classification
Preprocessing
• Point cloud data constructed from
RGB-D Kinect data.
• Discard points outside declared
outside the range of the sensor.
• Downsampled using a voxel grid
approach, with a voxel size of 1 cm.
Preprocessing
Candidate Leg Selection
• Use RANSAC to segment the floor plane in the point
cloud.
• Analysis of points limited to 20 cm above the floor
plane.
• DBSCAN ( Density-based spatial clustering of
applications with noise) clustering algorithm used to
identify foot candidates.
DBSCAN
• Three types of points
▫ core points
▫ reachable points
▫ outliers
• Two parameters
▫ Minimum number of
points required to form a
dense region (5 points).
▫ A radius to determine the
area to search for
neighboring points (4 cm).
Extension to Candidate Legs
• The foot candidates are expanded
to a height of 50cm using
DBSCAN.
• A Voronoi diagram is calculated
from centroids of the clusters
projected onto the floor plane.
• Expanded points are only valid if
their projections onto the floor
plane are in the correct Voronoi
region.
Classification
• Rectified depth image is computed for each
candidate by projecting its 3D points onto a
vertical plane of 40x50cm grid with a cell size of
1cm.
• HOG (Histogram of Oriented Gradients)
descriptor used to classify legs.
HOG descriptor
• The rectified depth image
is divided in 4x5 square
windows with an edge of
10cm, them discretize the
depth gradients within
each window over a 9-bin
histogram.
• This gives a 180
dimensional feature vector
that can be classified by the
pre-trained linear SVM.
Hardware (Kinect)
• 43°vertical by 57°
horizontal field of view
• Operating range of
~0.5-5.0 meters
• Depth and color images
resolution of 640x480
Experiments
•
•
•
•
•
Varying Height
Varying Lighting
Varying Range
Tested RGB and monochrome images.
Only indoors due to Kinect infrared sensor.
Questions?
Download