Documents
User Generated
Resources
Learning Center

# cricket

VIEWS: 84 PAGES: 20

• pg 1
```									Information Extraction from Cricket
Videos

Syed Ahsan Ishtiaque
Kumar Srijan
Problems Statement
• There can be many challenges like face detection
and recognition, boundary detection, separating
out the ad and the cricket part and many more.
• The problems we addressed are
–   Shot Transition Detection
–   Crowd Detection
–   Pitch Detection
–   Bowling Side Detection
–   Ball by Ball segmentation
–   Summary Making
Shot Transition Detection
• A shot is defined as a sequence of frames
captured by the camera in a contiguous way,
without interruptions.
• Shots are of two types
– Hard Transition
– Soft Transition
• Following are our solution to this problem
Normalized Cross Correlation
• Threshold the correlation between every 20th
frames
– By checking the correlation between every 20th frame, and if its
greater than certain threshold assuming that some transition has
occurred, then checking for correlation between every frame in that
interval, if any correlation value is greater than another threshold,
then Hard Transition else Soft Transition
• Results and Inferences
– The results were satisfactory for Hard Transition, but got false matches
for soft transition in the case when there was actually no transition
and the camera was moving very fast, as the correlation value for the
Kth and K+20th frame crossed the threshold.
– Also there was a problem that the correlation between the frames of
crowd was generally low, so generally every frame of crowd was
classified as a shot transition
Histogram Based
• Histogram based Detection
– By calculating the Histogram for all the bands or
for gray band, and thresholding for sharp change
of values we can detect for a transition.
• Results and Inferences
– This gave good results, the problem with this
approach was that it was not taking the position
of pixels into account
Hard Transition

Hard Transition

Hard Transition

Every frame here is a hard transition, but the Histogram
approach will not be able to detect this
Difference of NCC
• Difference of correlation values between
consecutive frames
– In this approach we checked the difference of correlation, and if
it is greater than a certain threshold we declare it as a hard
transition.
– Also blurring of the frames improved correlation and hence the
results
• Results and Inferences
– This also gave very good results, and by this approach the
problem of classification of crowd frames generally as shot was
also solved, as although the correlation between the frames of
crowd was low, but difference of correlation between the two
low correlation frame groups was not that high, so it solved the
problem.
Crowd Detection
• Detecting the scenes in which crowd is
present or is in focus.
Histogram Based
• Histogram based
– In a cricket video, we usually get two kinds of frames, one
with field where there are very narrow range of colors
present and of crowd where many different kinds of colors
are present.
– The histogram of crowd will be flat and will cover the
whole range, whereas other scenes will have histograms
which will be concentrated in a narrow range.
• Results and Inferences
– The results were not that encouraging. A single band or
combination of bands, and also the range of values within
a histogram could not be determined for histogram to be
constructed and analyzed.
Edge Based Detection
• Edge Based
– We observed that energy profile of the frame
containing crowd is distinctly higher than those of
not containing crowd, so we used canny edge
detector and build a edge map of all the frames,
now comparing the energy profiles of the edge
map solved our problem
• Results and Inferences
– The results were very good
Pitch Detection
• Finding whether the frame is showing a pitch
or not and determining its position in the
frame.
Template Based solution
• Template Matching
– We took a narrow horizontal strip, from the middle of a
frame showing the pitch, as the template. For each frame
in the video, we looked whether the strip can match
somewhere near the middle of the frame. We used
normalized correlation and square difference error as the
matching standards.
• Results and Inferences
– The results were not that good, had few false positives in
the frame showing the field.
– The problem with template matching was that it did not
address the variation of color of the pitch under different
lighting conditions.
HSI space based
• As template matching didn’t look into the variation of lighting
conditions
• The idea is that any frame not having the pitch will not be able to
match in all the planes, viz. Hue, Saturation and Intensity, even if we
take sufficiently large margins so that all the frames showing the
pitch are correctly classified
• For each frame, we broke it into HSI space and applied thresholding
on all the planes according to the ranges calculated before.
• We dilated each of the planes for filling the small holes and then
took the intersection of the resulting planes.
• We again dilated and eroded the resulting plane to remove the
holes.
• After that we took the distance transform of the image and
searched for the maximum value in that. Thresholding this
maximum value worked as the blob detection
Results
• The results were very good

Intensity             Hue    Saturation

Intersection           Blob    Original
Bowling Side Detection
• To detect from which side the bowler bowled.
• Solution
– Erosion of the blob obtained above shows the
skeleton of the pitch which is available.
– Whenever a bowler bowls the skeleton becomes
“L” shaped due to the occlusion caused by the
bowler.
– By detecting the orientation of “L” we can easily
determine the side from which the ball was being
bowled
Results
• This method gave very good results

Hue              Saturation   Intensity

Intersection        Skeleton      Original
Difficulties
• Later in the video as the part of the pitch was
covered with shadows, the “L” shape got
truncated and it was hard to determine the
bowling side.

Skeleton                   Original

Truncated “L”      Bowling side not
detected, but pitch
was still detected
Ball by Ball Segmentation
• To segment the video into deliveries
• Solution
– The detection of “L” can also help in knowing
when the ball is being bowled. When the skeleton
of the pitch changes from rectangle to L shaped
then a ball was bowled. We also made sure that
there was a gap of a few seconds between the
deliveries by setting a timer.
Results
Classified as a
delivery

This was not classified as a
delivery due to the timer

Again Classified as
delivery, as timer
got reset
Summary Making
• To generate a meaningful summary out of the video.
• Solution
– By a simple assumption that crowd is displayed only when
there is six, four, wicket or sometimes ads.
– So, every time crowd is detected we can go back few
seconds and take that part as highlight.
– Now sometimes crowd is displayed after an ad break, so
now pitch detection helps here, every time we go back we
can check for the presence of pitch, if it is detected then it
is not from the ad break and we can include that part in
our summary and if it is not we can skip that.

```
To top