Transcript Introduction - Computer Science Department

Mobile Model Car Navigation
Based on Wireless Video
Camera and Laser Sensors
Contents
Team introduction
Project Introduction
Applicability
The whole picture
Discussing tracking algorithms
A project development
Results
Team introduction
Supervisor:
Professor Ehud Rivlin .
Instructor:
Igor Katsman.
ISL Engineer: Yekutiel Pekar
Students :
Alexander Sherman
Valentin Rozental
Introduction
Mobile navigation on unknown
environment is an interesting and
challenging task.
Consists of several computer vision
based subtasks.
Each subtask is a project itself.
Where it could be used
Space Exploring
Medical applications
Military applications
Security applications
Civil Engineering
Science
…
Subtasks
Range-exploring algorithm for a mobile
model.
Tools
On-board wireless video camera
Laser sensors.
distance
Control interface between a mobile model and
a PC.
Tools
RC CAR remote control software module
The most interesting task
Automatic tracking algorithm for a
mobile model which would utilize
an on-board wireless video
camera to follow another leading
mobile model.
Tools
on-board wireless video camera.
General Scheme
server
frame
target
Turn Left
Lets focus on tracking algorithm
Tracking algorithm features
Simple
Its should be real-time algorithm
Robust
It should focus on mobile features recognition
Correct
It should minimize misclassifications, “loosing”
target, etc.
First steps
Matlab environment
acknowledgement
Edge detection
Contour recognition
Noises
Known Algorithms discussion
Temporal Differencing
Background Subtraction
Motion Regions Detection
Motion Region Classification
Velocity Vector Computation
Advantages
Simple
Robust for static camera tracking
Background subtraction eliminates all static
background’s pixels and moving objects could be
easily detected.
No information is needed about a starting
position
BUT …
Disadvantage
A camera is in significant motion all the time.
NO STATIC BACKGROUND presents.
Solution
Image stabilization algorithm
Drawbacks
It computationally expensive
Hard to get real-time speed
Conclusion
Template Correlation matching algorithm may be
used
Another question :
What is good template feature ???
Contour
Advantages:
depends lightly on light conditions
good for the simple objects
Disadvantages:
Hard to detect if the noise level is high
Complex forms in different visual angles
A big variety of such forms
Color
Advantages:
avoid the noise problem
more effective use of filters
Don’t depend on object form
Easier to detect the object location
center of mass
Disadvantages:
depends heavily on light conditions
Color is a feature that is widely distributed
Size
Advantages :
Easy to detect if the object is big
Doesn’t depend on light conditions
• Disadvantages:
– Easy to misclassify the object if the object is small
– Very sensitive to angle of sight
– Depends heavily on the object distance
Template
Advantages:
Easy to scan the image
Could be done by partial match
Less sensible to noises
Disadvantages:
A variety of templates must be provided and
(sequent/parallel) diagnosed
Not a real-time method
NO “SILVER BULLET” !
But still there are good choices !
A choice
Simplicity is desirable
Accuracy is mandatory
Real-time performance is crucial
Color Histogram
As flexible as memory-based methods
more compact representation
Estimation is trivial
Color histograms for an image are built from pixel
values in one of color spaces.
Different weight to different color ranges
increases probability of quick and precise target
detection.
Algorithm
When new image becomes available
A local search is performed
The best match is found
Velocity vector is calculated
Matching
For each searched area
Get a color histogram
Normalize it.
Find an intersection with previously stored template histogram
minimize differences
Get a color histogram
color _ range _ vector (i) 
ranges
 color _ range _ vector ( j) * weight _ vector ( j)
j 1
Histogram 
N
 (low _ value  color (i) high _ value)
j 1
Normalize and compare
Histogram
normalized _ Histogram 
number _ of _ pixels
result 
ranges
 | template _ histogram(i)  normalized _ Histogram(i) |
i 1
If result < best result
best result = result
Making the tracking more robust and
efficient
Template updating
Template history management
Motion area prediction
Template updating
Solves disadvantage of great sensibility to light
conditions
Ensures that current template accurately
represents the new image of object
A method
merging the previous instance of template with
current information we have
Template History Management
Solves partially or full occlusion of the object
Prevents from misclassification and “loosing”
the target
A method
merging the previous instance of template with
current one
A history parameter is estimated empirically
For example:
R(n) = a * M(n) + (1 - a) * R(n – 1)
where
R(n) – is a template at n-th stage
M(n) – is a new motion detected
a
– is a history parameter
Motion Area Prediction
Processing a whole picture is
computationally very expensive and is
inapplicable to real-time applications
or requires specialized hardware to operate in
the real-time domain.
To overcome this obstacle we use a motion
region prediction that reduces search area.
A Method
The 2D image velocity vector of the
target (u, v) (pixels/frame) is approximately
determined
Calculating the difference between centroid of
the previous template R(n-1) and centroid of
the new template R(n)
In the area of centroid exhaustive search is
performed. In case this search fails then area is
enlarged in such way that ensures object
detection.
First results
First we tried to track simple movements in
slow motions
Back and Forward
motions
Diagonal motions
Car come across
Parallel motion
Car is approaching
Car is moves away
Making things more complicated
Then we decided to add some obstacles to
see how we deal with them
Car is partially
occluded
Car is occluded by the
obstacle of the same
color range
Final Exam
 freestyle 
A pleasant surprise
Although we
concentrated on car
tracking it still capable
of tracking other
objects
A HELICOPTER
AND finally some FUN
See
anybody
you
know?

Conclusions:
Using template correlation matching has
three main advantages:
Continuous tracking despite occlusions and
cessation of target motion.
Prevents template drifting onto background
texture.
Provides robust tracking.
Although a module show good results it
has several drawbacks. Some of them are:
Hard to detect if the noise level is high.
The probability to misclassifications of the
object is great, since the color is a feature
that is widely distributed.
Suggestions:
To overcome these drawbacks gradientbased algorithm might be implemented as
secondary stage recognition. This will
strongly decrease an probability for
misclassification and “loosing” a target.