Use the visual feature extraction technique to identify football players in keyframes of a video. Also, apply the similarity matrix to analyze the variation and movement of the match.

1. Hat Trick Coach
pchi4616 - 490250497
dpan7802 - 480180371

2. Motivation
 Problems
 High domination does not mean winning game. How they act in key events?
 Throughout the match determine the different patterns the teams follow and play
in the match
 Determine the different scenarios which may arise during the match(such as key
events).
 What did the winning team do in order to win and how the losing team lost the
edge in the game.
Football Field Ontology of a Football Match
Corner
Sideline
Goal Box
Penalty Box
Center Circle
Goal Line
Object:
 People: Goalkeeper, Defender,
Midfielder, Striker, Coach,
Referee, Audience
 Facility: Ball, Goal Post, Grass,
Line
Event: Goal, Corner Kick, Foul,
Injury, Running, Special Skills
Concept: Condensed Press, Sparse
Formation, Win, Lose and so on.
Corner
Sideline
Goal Box
Penalty Box
Center
Circle
Goal Line

3. Motivation

4. Approaches Video Uploaded
Scene Detection
Morphological
Transformations
Color Masking
Statistical Feature
Extraction
Video Feature
comparison
Show statistics results
Image Feature extraction
Cosine for similarity analysis

5. Scene Detection
By monitoring variation of HSV color space, extract a sequence of scenes.
Based on the result, we know in a football match, scenes appear whenever
there is a zoom-in or zoom-out.

6. Analysis of Frames in Scenes
 For each scene, we extract a frame every second, and put extract features
from each frame.
Selected Frames Not selected frames

7. Image Feature
Extraction
To identify players, we need to extract 2 image
features:
 Shape Feature: Masking the field and
conduct the morphological operation.
(contour)
 Color Feature: With contours, we set a
filtering condition to the pattern of shape
and use masking operation to identify
players’ teams.
 h >= (1.2) * w and h>10 and w >10
 Number of colored pixel > 30
frame masking
Morphological Transformation Player identification
Masking of the contour

8. Statistical Feature
Extraction
 Number of Players
 Distance of Players
 Distance Metric: 𝑛𝑜𝑟𝑚2
 Normalization of Distance Matrix:
 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 =
(𝑥1−𝑥2)2+(𝑦1−𝑦2)2
max _𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒

9. Result

11. Future Work
 Consultation with the Coach
 Improve the recognition system
(when player are grouped together or
overlapping they are hard to identify)
Coach’s Feedback:
A football field can be
separated into several
areas, it would be more
helpful to identify the
exact area where each
player locate. Also, it’s
useful to identify the
shooting area and offside
area

12. Reference
 1. Baraldi, L., C. Grana, and R. Cucchiara, A deep siamese network for scene detection in
broadcast videos. arXiv preprint arXiv:1510.08893, 2015.
 2. Datta, S. and N. Oschlag-Michael, Understanding and Managing it Outsourcing: A
Partnership Approach. 2015, Basingstoke: Palgrave Pivot.
 3. Kapela, R., K. McGuinness, and N.E. O’Connor, Real-time field sports scene classification
using colour and frequency space decompositions. Journal of real-time image processing, 2017.
13(4): p. 725-737.
 4. Abdullah et al. (2018), “Soccer Event Detection,” Computer Science & Information
Technology (CS & IT), pp. 119-129.
 5. Adrian Rosebrock (2018), “Yolo Object Detection with OpenCV,” Deeping Learning,
Objection, Tutorials, pyimagesearch
 6. Brandon Castellano, PySceneDetect Document,
https://pyscenedetect.readthedocs.io/en/latest/, 2018.
 7. KananVyas, Player Detection, Github Project:
https://github.com/KananVyas/PlayerDetection, 2018.

13. Thank You For Your Listening