3. introduction
• Proposed a appearance based representation of
target based on kernels and color histograms
• Input region of the target object in the previous
frame and stored instance of target object and
tries to localize it in current frame
• For background subtraction similarity of color
histogram we use
• Since the object view is changes with respect to
its motion, we are employing the comparison of
regions of interest in the current frame and
previous one
5. MODEL
BASED
Priory
informatio
n about
the object
Problem in
tracking
under
illumination
variation
APPEA
RANCE
BASED
Object
projection
on the
image
Sensitive
to
illuminati
on
CONTO
UR
BASED
Shape
matching
or contour
evaluation
technique
Enables
tracking rigid
and non
rigid object
FEATURE
BASED
Tracking
according to
the features
Perform well in
partial occluded
situation
Problem in
tracking object
and background
HYBRID
BASED
Employs
the above
two
technique
s
Disdv.high
complexit
y
7. LSK object tracking
• This appearance based one for tracking both
rigid and deformable object without the prior
information of object
• Assumption is that the object translation and
deformation between consecutive video is
small
• Searching for ROI in the video frame from a
search region
8. To track small object appearance change
Cosine transform of LSK Prefers maximum LSK
The similarity of the LSK and CH
Between ROI and object region
Btw the previous frame and object
instant
From new frame ,the region of interest we take
Spatial information through LSK
Color information through color
histogram
9. Initialization of object ROI
in first frame
Color similarity search
in current search
region
Representation of
object and selected
region through silent
feature
Decision of object
ROI in new
frame
Update the object
instant in the stack
Prediction of object in
the video frame
10. Color similarity
• It is discriminate the object from background
by color histogram comparison
• but it is sensitive to illumination
• Suppose search region is R1×R1 is divided into
ROIs , shifted by d pixels vertically and
horizontally and having size equal to size of
the query object Q1×Q1
11. • Total patches R1-Q1+(R2×1/d)-Q2+1/d
• 1/d density of the uniformly selected ROI
• Increase in d a uniform sampling of candidate object
ROI every d pixels in the search region
• Minimal histogram similarity indicates background
• The CHs are compared according to the cosine
similarity or some matrices ,Bhattacharya distance
• Let h1 and h2 be the two histograms and the
comparison of two is
12. •
• In order to map the range [-1,1] to
we apply
• This is calculated for 3 color channels
• Highest CH similarity is belongs to object
• The similarity value of all possible patches
comprises a matrix
13. • Patches with lower CH is background and
higher CH is object
• A threshold value is for each frame
• Confidence level Bt%
• This decreases where background color is
closer to object color
14. Object texture description
• Edges carry more information about the
object
• Texture description is given by LSK, which
exploit both spatial and pixel value
information
• The distance between image pixel p and it
neighboring pixel pi which employs the
covariance matrix ci
15. • Ci is the gradient containing information about the
dominant edges described by eigen value.
• To calculate the ci,we use
• matrix
• F(p)calculates correlation matrix
16. • Extract the LSK vector
• And normalize It ,invariant to
illumination changes
• LSK are good image texture descriptors,
because they are invariant to brightness
changes, contrast changes and noise.
• Here we are using lab color space
17. Object localization and model update
• Divide the object search region in to overlapping
patches of equal size
• We extract CH and LSK
• For each patches ,check the 3 cosine similarity
• This is to avoid tracker drifting
• Final decision matrix is
• are LSK similarity matrix and is binary
CH similarity matrix
18. • To detect the object by±ϕ degree, we rotate the
video frame t, around the center of predicted
object position
• Two decision matrix calculated according to the
previous equation:
• zooming can be detected by rescaling the search
region ±s%
• Here also LSK similarity is checking
• Experimental value is 10 degree and 10%
19. SEARCH REGION EXTRACTION IN NEXT
FRAME
• This is to predict the position of object in the following frame
• By kalman’s filter we predict the position of object in current frame
• Noise covariance matrix
• The linear kalman’s filter is efficient method for motion state prediction
20. Experimental result
• Initialization perform through a training free
object detection algorithm
• The search region is R1×R2=2Q1×2Q2
where Q1×Q2 is downscaled object dimension
Window size of LSK is 3×3 and rotation in 1D is
10 and scale step is 10%
Threshold is
21. • DONE IN TWO WAYS
• Appearance base object representation with
particle filter
• Object tracking by dividing the object of
interest in smaller fragments
• For better evaluation PF tracker and FT tracker
initializes with different ROIs
22.
23. • In first case we track the object bag,2nd we are tracking the
car in the video.
• FT tracker does not take into account change in scale and
CH-LSK tracker is successfully in tracking in the change in
object image size but PF tracker tracks constant size.bt in
2nd experiment both has similar performance
24. • In 3rd case we track a car in trajectory in Omni
directional camera
• FT tracker is not use ,because it does not handle
rotational motion
• We are using proposed algorithm and PF tracker
• PF tracker is loses the object quickly and CH-LSK
follows better for rotation
25.
26. • In 4th and 5th case we tracks object in partially
occluded and a small scale variation .
• In this we track a man is coming towards the
camera
27. • CH-LSK tracks the man’s torso, contain significant
background data
• In 2nd experiment , we are tracking a man
• Here FT filters are used
• PF tracks the man but FT cant well
28. Case 6th :tracks the
video with strong
illumination
FT cant track the
illumination
variation WHILE
CH-LSK tracks with
the illumination
variation
Visual object
tracking can be
employed to track
human activity by
analyzing the
utensils that they
use
By tracking
distance between
the utensils and
the face we track
motion
29.
30. Case 7
• The PF tracker cant track the motion of the glass and loses while sipping
• FT loses while glass moves up and down
• CH-LSK tracks succesfully
case8
• Hand tracking is difficult but PF and ft can track right hand but not left
hand
• Proposed one tracks it successfully
Case 9
• Glass is rotated and changing shape occluded
• Proposed one tracks successfully
31.
32. • Case 10:tracks constantly changing orientation
and hands occlude part of face
33. QUANTITATIVE EVALUATION
We are evaluating this by frame detection
accuracy
• G is the overlap area between ground truth obbject
Nt is the no of object frame
Range from 0 to 1
36. conclusion
• CH-LSK has the good capability to track the
object under several changes in appearance
and partial occlusion.
• limitation: does not handle full occlusion
case
• it tracks another object in background also
• Kalman’s filter cant follow sudden changes in
the object direction and speed, decreases the
algorithm speed