Transcript lateckiResearch08

Computer Vision and Data Mining
Research Projects
Longin Jan Latecki
Computer and Information Sciences
Dept.
Temple University
[email protected]
Research Projects
• Object detection and recognition in images
• Improving ranking of search queries
• Motion and activity detection in videos
• Merging laser range maps of multiple robots
Object detection and recognition based on
contour parts
• Often only parts of objects are visible in images
• We can detect and recognize such objects in
edge images by performing contour grouping
with shape similarity
Edge image
Detected object
Algorithmic overview
Methodology
• Probabilistic approaches are needed to address
noisy sensor information in robot perception.
• We use Rao-Blackwellized particle filtering that has
been successfully applied to solve the robot
mapping problem (SLAM).
• We use medial axis (skeleton) as our shape
representation.
•Supported by DOE, NNSA, NA-22
•NSF, Computer Vision Program
Sample evolution of particles
Iteration 2
Iteration 14
Iteration 10
Iteration 18
Experimental results
Bottle model
Swan model
Reference models
Bird model
Applications:
Analysis of aerial and satellite images,
in particular object and change detection
Supported by LANL, RADIUS: Rapid Automated Decomposition of
Images for Ubiquitous Sensing, PI: Lakshman Prasad, LANL
the original aerial image
detected parts of contours
detected structures of interest at three different scales (in maroon).
Motion and activity detection in videos
Object and activity
detection results
Videos are obtained from the Temple University Police video surveillance system.
Methodology:
We use PCA to learn local background textures, and detect motion by analysis
of texture trajectories.
Many Video Surveillance Applications, e.g.,:
Detection of moving objects and detection of abandon objects,
e.g., around power plants
Human detection in infrared images and videos
Improving ranking for similarity queries
original
improved
query
original
improved
Improving ranking in face profile retrieval
Original retrieval
query
Improved retrieval
Methodology:
We use semi-supervised manifold learning to learn new distances
in the manifold spanned by the training data set.
Further applications:
This methods makes it possible to improve ranking of any queries
from images through text to concepts.
Prior based
model
Merging
mapson
ofmotion
multiple
robots
• Our motion model is based on structure registration process
between local maps which results in multi-modal prior.
Prior in odometry based
motion model
Prior in our structure
registration based motion
model
Experimental results
Dataset: NIST Maze data set
Merged global map
Sample individual local maps