Irish Interdisciplinary Journal of Science & Research (IIJSR)

Vol.4, Iss.3, Pages 66-78, July-September 2020

Prediction and Classification of Fouls in Soccer Game using Deep Learning

T.Thamaraimanalan1, D.Naveena2, M.Ramya3 & M.Madhubala4
1
Assistant professor, Sri Eshwar College of Engineering, Coimbatore. Email: thamaraimanalan.t@[Link]
2
Sri Eshwar College of Engineering, Coimbatore, India. Email: naveenadurairaj1999@[Link]
3
Sri Eshwar College of Engineering, Coimbatore, India. Email: ramyaraghav372919@[Link]
4
Sri Eshwar College of Engineering, Coimbatore, India. Email: madhubalam969@[Link]

Article Received: 11 June 2020 Article Accepted: 09 August 2020 Article Published: 06 September 2020

ABSTRACT

Background: Event detection within the games in the challenging task in the field of object detection system. Event of the certain games represents the various
foul factors, player action as well as the object tracking. The proposed work implies the tracking of the football which is followed by the players tracking that
implies in the action of the event within the football game. Various games can be analyzed using the deep learning methods which yield the discovery of fouls and
penalties in certain games in Olympics. This can lead towards change in results in games with the accurate prediction.

Method: The object detection within the football games depends upon the basic image processing techniques. The image acquisition stage involves the collection
of database as in the video format or in the live data stream. The pre-processing stage is carried out using the background subtraction technique using the binary
conversion technique. Histogram optical flow and Local Binary pattern is carried out for the segmentation of movements of players as well as ball tracking.
Features are extracted and stored from the segmentation process. Training process is carried out using Convolutional Neural Network (CNN) classifier.

Results: The trained and the tested phase of the proposed work maintain the accuracy rate of 87.63% with the reduction in the sensitivity and the specificity range
of 72.5 and 86.2%. The experimental results are carried out using the Python domain which uses the tensorflow platform that yields the lesser error rate and
higher accuracy of classification.

Keywords: Image acquisition, Background subtraction technique, Optical flow and Local binary pattern.

1. Introduction

In our daily lives, computers are an integral part. Although it is productivity, the web is accessed for
gathering data, or even for playing games. In the past few periods, computer growth had contributed to
great strides and development in fields such as Mathematics, Finance, and Biology. Meanwhile the arrival
of Machine learning, computers have been able to process further like the quality of humans such as image,
form, speech, and even in visual object recognition [19]. The accessibility of a huge quantity of information
in Machine learning is the key reason for success. In this field, the research had been prosperous due to the
recent abundance of information. So that it contributes towards algorithmic and computing performance.
The machine had been allowed to examine and study from this data.

Computer Vision is one of the most fascinating areas of Machine Learning. It includes the derivation of a
high-level understanding of low-level raw pixel data [40]. To study this field is more difficult due to the
complication in an understanding of what these pixels represent. For example: Given a picture of a sunset
and asked to explain the image, a human, below usual circumstances will be capable of quickly recognizing
the sunset. A machine, though, is not quite so talented. The raw pixel data is processed by the image but the
machine does not know what this data signifies. Developments in Computer Vision helps to bridge the gap
between the machine's and the human brain's abilities to interpret visual signs. Computer Vision has
recently been in advance power. Image classification and object detection are the subdivision of Computer
Vision which is being changed to fields such as engineering and medicine. In the automotive industry, the
study has been proposed using both for an application like autonomous cars with favor to pedestrian
detection [42] and road detection [25]. Video games in Computer Vision are another interesting area. A
study published in 2015 proposed that with the current beginning of the e-sports industry and growth of

ISSN: 2582-3981 [Link]

66
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
inexpensive gaming, live streaming of sports will rise by 140.6% among 2014 – 2019 [29]. The study also
forecasts the growth concerning the viewing audience for sports around 305.5% within the same period
[29]. This increased evolution of competitive tournaments opens new paths for Computer Vision
exploration.

Examples of possible benefits of this type of examination include: gathering information on player habits to
find faults or enhancements, development of more complex game AI to assist players to learn faster, or
executing ideal stability updates. Computer Vision research in sports is presently lacking. There are few
studies which seem to attention specifically on image classification combined with reinforcement learning
or Q-learning and producing agents that can play the game [18], [26]. Further subdivisions of Computer
Vision such as object detection don't seem to have been familiar by the gaming community as a useful area
of study. By considering this, we decided to push the partial examination done in this area by applying
object detection to video games.

For our study, we chose to concentrate on object detection as a total with numerous core classification
perceptions. We applied a real-time detection model to organize and track video game characters from the
popular fighting game Super Smash Brothers Melee. By this, we constructed a basic bot, capable of
measure based on tracked locations of a secondary character on screen. This bot was accomplished using
labeled frame data from the game and successfully achieve basic actions in real-time.

2. Related Works

Till the detection of deep learning, sports video observation, particularly in football video analysis, had
been classified into two group pattern recognition [21, 9] and object tracking. In case of object tracking
personalized camera was used results in estimation cost, however the low-level events were extracting by
pattern recognition and then high-level events were detected by the classifier.

Qian et al [17] proposed a limited method, that is used with clear success for soccer movement recognition
contains classification of events into a discrete category like the goal, shoot, and so on. This approach
contains heuristic rules and feature extraction for spotting actions. They accomplish simple event analysis
to detect symbols (arcs, goalmouth, fields, lines, and logo), ball and player position etc., and it originates
between complex event and simple event features using the signs.

Finally, they built up a standard based framework to distinguish applicable events like the corner, the goal,
and so on. Jin et al [10] proposed a Hidden Markov-based calculation for detection of video event
dependent on signs blend and mix. In soccer video analysis, the detection higher-level from low-level
events is highly difficult problem.

The exposure reveals, e.g., the movements of the ball and the player on the pitch, which might be useful to
classify definite movements ("passing the ball", “shot on goal”, etc.) or too well known the overall trend of
the game. Later in 2012, a deep learning approach such as Restricted Boltzmann Machines and Convolution
Neural Networks have been efficiently used for activity and event detection. Convolution Neural Networks

ISSN: 2582-3981 [Link]

67
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
have exposed superior performance in object detection, modeling high-level and image classification
pictorial semantics [6], [11], [8]; Temporal dynamics video is exhibiting by the recurrent neural network
[12].

Normally utilized activity confinement procedures, for example, fast r-Convolution Neural Network and
faster r-Convolution Neural Network [18], [7], for the most part start with the area of interest (proposition
age) to get a lot of up-and-comer locales, around then utilize a completely related layer toward the
conclusion to classifying goals.

Recent approaches declared above concentrate on recognition of events in football game from the
perception of classifiers, models, and feature extraction for extracting simple events. Such methods lack in
meaningful representation of in between events. Introducing semantic explanation and essential
information in these approaches is quite problematic. Thus, this urges us to begin from the essential
structure squares and recreate a framework that grants abusing the semantic information around event,
which can be used to distinguish the high-level and intermediate type of events.

As far as we could possibly know, however there are frameworks that consequently recognize essential
evidences, correspondingly the position and movements of the player, there are no programmed pointers for
semantically composite event, such as scoring on a corner kick or penalty kick.

3. Proposed Work

3.1 Video Events

An exact ontological description of the event is an open fact. The idea of this section is to define our event
detection in precise manner. There are so many events were conducting but the video events is the most
interesting one. For example, a football game is video event. In this “Pass the ball” is one of the events in
which the distance among player1 and the ball rises and the distance among the player2 and the ball fall.

3.1.1 Low-Level Events

A low-level event type is defined as follows:

𝑆𝐸 (𝐼, 𝑠𝑖𝑚𝑝𝑙𝑒𝐸𝑣𝑒𝑛𝑡𝑇𝑦𝑝𝑒, 𝑡, (𝑟𝑜𝑙𝑒1 , 𝑜𝑇𝑦𝑝𝑒𝑖 ), … . , (𝑟𝑜𝑙𝑒𝑛 , 𝑜𝑇𝑦𝑝𝑒𝑛 )) (1)

Where, I is the attribute which represent the event, simpleEventType is the type of event, e.g. "throwing the
ball", and t is the duration of time within which the event take place, role1..., rolen (n = 1, nmax) are the task
that dissimilar objects play in type of event. For example, “throwing the ball” is a subject who throws the
ball is first task and second task is thrown ball; at last oTypei is the legal type of object that can play the task
rolei, e.g., it is only balls can be thrown and only players who can throw. Examine, the entire definition of
the type of event "throwing the ball" is

(𝐼, 𝑇ℎ𝑟𝑜𝑤𝑖𝑛𝑔_𝑡ℎ𝑒_𝑏𝑎𝑙𝑙, 𝑡, (𝑡ℎ𝑟𝑜𝑤𝑖𝑛𝑔𝑃𝑙𝑎𝑦𝑒𝑟 , 𝑝𝑙𝑎𝑦𝑒𝑟), (𝑡ℎ𝑟𝑜𝑤𝑖𝑛𝑔_𝑜𝑏𝑗𝑒𝑐𝑡, 𝑏𝑎𝑙𝑙)) (2)

A particular instance of a low-level type of event defined in (1) the following tuple:

ISSN: 2582-3981 [Link]

68
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
(𝐼, 𝑠𝑒𝑇𝑦𝑝𝑒, 𝑡, (𝑟𝑜𝑙𝑒1 , 𝑂1 ), … . , (𝑟𝑜𝑙𝑒𝑛 , 𝑂𝑛 ) (3)

Where, I is the attribute which represent the event, O1 and On are objects identifier observed in the frame
related with the duration of time t, correspondingly.

(12, 𝑇ℎ𝑟𝑜𝑤𝑖𝑛𝑔_𝑡ℎ𝑒_𝑏𝑎𝑙𝑙, 𝑡, (𝑡ℎ𝑟𝑜𝑤𝑖𝑛𝑔_ 𝑃𝑙𝑎𝑦𝑒𝑟, 𝑜𝑏𝑗02), 𝑡ℎ𝑟𝑜𝑤𝑒𝑑_𝑂𝑏𝑗𝑒𝑐𝑡, 𝑜𝑏𝑗01)) (4)

The event of "Throwing_the_ball" defines a low-level type of event "throwing the ball" that occurred at
time duration t, anywhere the obj01 is thrown by obj02. Also, obj01 and obj02 are two objects in the
structure were detected corresponding to duration of time t, of type player and ball in that order.

3.1.2 High-Level Events

High-level events are built for the purpose of more accuracy. High-level events are detected by the simple
events. We can use temporal and logical operators to detect high-level event. The hierarchy events are
definite by low-level event to high-level event. In the following we have two types of complex event one is
logical event and another one is temporal event.

3.1.3 Logical events

A logical event originates from the function of logical operators such as NOT, OR, AND to a set of events
which might be complex or simple.

𝐿𝐶𝐸 = (𝐼, 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝐸𝑣𝑒𝑛𝑡𝑇𝑦𝑝𝑒, 𝑡, 𝐿 =< 𝑒1 𝑜𝑝 𝑒2 𝑜𝑝 … . . 𝑜𝑝𝑒𝑛 >) (5)

Where, I is the attribute which represent the event, complexEventType is the high-level type of event (such
as "The goal is valid only if there is no foul"), t is the duration of time in which the high-level type of event
exists, L is the lower-level set of events e1. . . . . . en joined by op logical operators (i.e. NOT, AND, OR).

3.1.4 Temporal events

A temporal event stems from the application of temporal operation than as follows:

𝑇𝐶𝐸(𝐼, 𝑐𝑜𝑚𝑝𝑙𝑒𝑥𝐸𝑣𝑒𝑛𝑡𝑇𝑦𝑝𝑒, 𝑡, 𝐿 =< 𝑒1 𝑇𝐻𝐸𝑁 𝑒2 … . 𝑇𝐻𝐸𝑁 𝑒𝑛 >) (6)

Where, I is the attribute which represent the event, complexEventType is the high-level type of event (such
as “player 1 passes the ball to player 2"), t is the duration of time, L is the arrangement of simple events,e1,en
that should follow in the order. For precedent,e1, e2, e3, e4 may be, correspondingly, “player1 possesses the
ball", “player1 kicks the ball", “the ball approaches player 2", “player2 gets in possession of the ball".

4. Event Types

The mainly important factors in analyzing a soccer game is the capacity to identify the goal, kicks, passes,
offside, cards, ball possession and many other events in a video. Till date the event recognition in Soccer
field is done by using immovable cameras to capture the movements of players and the ball. By using this
method, the accuracy of movement tracking in the soccer field has enhanced to greater levels but this
method is too costly. The video clip that is used can also be accessed from Internet. Here we are explaining

ISSN: 2582-3981 [Link]

69
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
a tiny event which is the most complex event in soccer game which are in continuous frames, for instance
the ball possession and the player stroking the ball found on the distance among the bounding boxes and the
rules which is a mix of logical and temporal operations is set for each and every event.

In this process, the event is defined by a set of rules and regulations, but this will not be completely flexible
and on further up gradation of this event detection may have a possibility of detecting the events
automatically by using deep learning techniques.

4.1 Ball possession Event

The two terms TBP and PBP are combined known as ball possession. TBP means team ball possession and
PBP means player ball possession. These two possessions have same starting point but different ending
point. When the ball is in play it determined the ball possession by the intervals of time. Player ball
possession starts when a player starts to kick the ball at initial point and it ends when the ball halt in the
field.

PBP can be said as “An event takes place when the space between the ball and player is not beyond the
threshold value and the ball is close to that player”.

(𝐼, 𝑃𝑙𝑎𝑦𝑒𝑟𝐵𝑎𝑙𝑙𝑃𝑜𝑠𝑠𝑒𝑠𝑠𝑖𝑜𝑛, 𝑡 + 𝑘̅ , (𝑃𝑜𝑠𝑠𝑃𝑙𝑎𝑦𝑒𝑟, 𝑝𝑖 ), (𝑃𝑜𝑠𝑠𝑂𝑏𝑗𝑒𝑐𝑡, 𝑏)) ←

𝑝𝑙𝑎𝑦𝑒𝑟(𝑝𝑖 ), 𝑏𝑎𝑙𝑙(𝑏), 𝐷(𝑝𝑖 , 𝑏, 𝑡) < 𝑇ℎ ,

∀𝑗 ≠ 𝑖, 𝑝𝑙𝑎𝑦𝑒𝑟(𝑝𝑗 ), 𝐷(𝑝𝑗 , 𝑏, 𝑡) > 𝐷(𝑝𝑖 , 𝑏, 𝑡)^

∀𝑘 = 1 … . 𝑘̅ , 𝐷(𝑝𝑖 , 𝑏, 𝑡 + 𝑘) ≈ 0


The PBP takes place at the time interval t  k and when the distance D between the ball b and the player pi
at time t(pi, b, t) is below the threshold Th and D(pj, b, t) is the distance among any one player pj and the ball
b is more than D(pi, b, t). Also, the communication between the players and the ball is not always the same;
it will be less after an action up to certain number of continuous frames. The threshold value T h, is an
experimental calculation of the physical interaction among the player and the ball.

4.2 Kicking the ball Event

In a video clip of soccer which is a continuous arrangement of frames, the type ofevent regarding to the
Identifying the ball after kicking the ball, and the distance among the player and the ball vary continuously.
At the beginning, the distance among the player and the ball is less and then later the distance among the
player and the ball is more for certain number of frames and after certain period of time the interaction
between the player and the ball can no longer happen and so we can define the type of event “kicking the
ball” as follows:

(𝐼, 𝐾𝑖𝑐𝑘𝑖𝑛𝑔𝑇ℎ𝑒𝐵𝑎𝑙𝑙, 𝑡 + 𝑘̅ , (𝐾𝑖𝑐𝑘𝑖𝑛𝑔𝑃𝑙𝑎𝑦𝑒𝑟, 𝑝𝑖 ), (𝐾𝑖𝑐𝑘𝑒𝑑𝑂𝑏𝑗𝑒𝑐𝑡, 𝑏)) ←

𝑝𝑙𝑎𝑦𝑒𝑟(𝑝𝑖 ), 𝑏𝑎𝑙𝑙(𝑏), 𝐷(𝑝𝑖 , 𝑏, 𝑡) < 𝑇ℎ ^

ISSN: 2582-3981 [Link]

70
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
∀𝑘 = 0 … 𝑘̅ − 1, 𝐷(𝑝𝑖 , 𝑏, 𝑡 + 𝑘) < 𝐷(𝑝𝑖 , 𝑏, 𝑡 + 𝑘 + 1)

The above expression can hold as long as the distance among the player and the ball raises after the player
interact with the ball. In soccer game there are many types of kicks. They are Cornor kick, Penalty kick,
Goal kick, Free Kick, and so on.

4.3 Restrictions of the proposed work

When we evaluating a game, there can be many fouls happening throughout the match but we are
considering only the major fouls. In certain cases, the contact between the player and the ball may not be
there but the player runs nearing the ball to reach the ball. The interaction among the player and the ball
happens throughout the match. In other cases, when the player and the ball are near but the ball is behind the
player i.e. the player is standing showing his backside to the ball. For a superior understanding among the
different types of kicks, we can identify it by the velocity at which the ball travels after the interaction
among the ball and the player. After dribbling a ball, it will be slower when compared to the ball after a
kick/shoot. We too taking into account of the threshold level of every players by considering about each and
every players profile linked to their usual communication with the ball.

5. Methodology

The block diagram of the proposed architecture shows the method of work. The data consist of
approximately three minutes long video, which consists of about 6200 frames. Using SSD players and the
soccer ball are detected from each single frame. After that unwanted parts of the frames in the input were
filtered out because they are not defined about events. In filtering process they keep the data around the
bounding boxes using logical and temporal operators as a result we detected the fouls.

Fig.1 Block diagram of the proposed architecture

FRAMES: Frames are nothing but it is a sequence of events running as video. Each and every frame
consists of bounding box set. It represents the dimension and position of the object. To detect the object first
the sequence of frames were given as input to the SSD and coordinates of the object were known by the
bounding boxes with confident score.

SSD: SSD stands for “Single Shot Multi-Box Detector”. In early application one method is common for
object detection that is the system provides a part of region which re-samples the pixels and bounding box
provided with features followed by classifier. It detects the event but it is very expansive for low frame rate
in computation process. There are another method in detection of object is High-speed single shot
multi-box detector. It contains two main tasks one is classifier and another one is region proposal. The main

ISSN: 2582-3981 [Link]

71
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
plan behind single shot multi-box detector is s a narrow convolution filter is applied to meet the bounding
boxes.

Fig.2 Performance accuracy plots

SSD requires an image that should be supplied as input and the information about the ground for each
object class while training the frames. For training the images, first the models of a real soccer match video
are required. Vatic Animation tool allows annotating of players and ball inside each frame drawing a
bounding box around the players. The outcome of this method is a group of pictures with coordinates of
bounding boxes stored in Pascal VOC layout. The performance accuracy plot is shown in fig.2 and the
comparison of existing algorithm is shown in fig.3. Table 1 shows the manually annoted numbers of an
object, and it is used for the training and test of single shot multi-box detector.

Fig.3 Comparisons with existing algorithms

Table 1. Manually Annotated Object to Train and Test the SSD

Player
Player Ball Flag
Movements

Training 1235 529 125 15001

Test 867 318 76 11440

Result 2102 847 201 26441

ISSN: 2582-3981 [Link]

72
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
In Table 1 the training set has been used to form a SSD representation. In the below Table 2 the test set of
average precision value is given. For instance from the soccer match the input and the output appearance to
SSD is made known in Figure 4 and Figure 5, correspondingly.

Filtration: Filtration can be done by the particular threshold value of the object detected from the single
shot multi-box detector. For example, in a single frame multiple players were detected but we required only
a particular threshold values which defines about the events other than that we can filter out.

Table 2. Average Precision of the System

FP FN

TP 2 2

TN 189 150

Detection of Event: In a real-time application, we have some detection regarding video events. In that case
there was sequence of frames running over a time period. We cannot study about all the frames running
behind the video clips and it is highly complex. This methodology is going to have a solution for problem.
The system can detect the events by this process and there are two main steps have been done. One is the
detection and another one is a collection of simple low-level events. By keeping this as a base we can also
detected the highly-complex events. The system also detects the type of event.

We can also feed the new upcoming events with unique event type identifier. The events can be monitored
by two levels, high-level events and simple low-level events. The simple low-level events recognize the
simple event and high-level events recognize the complex events.

The detection of event by the system revives bounding boxes with confident score as input initially it
process the input to identify which type of events based on the rules and it stores the event in the memory.
The input bounding boxes nothing but it defines about the object coordinates. After the event detection we
use temporal and logical operators to know the event is high-level event. There are so many programming
languages are there for event detection but we have taken python language because it is general purpose
syntax.

Fig.4 Original frame

ISSN: 2582-3981 [Link]

73
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020

Fig.5 Detection of object by SSD with confidence score

6. Experimental Results

The proposed systems are used to detect the low-level complex occurrence in the real football video such as
kicking the ball and ball possession. The work has been tested by 3 minutes video consists of approximately
6.2k frames. While experiment that process we know that comparing to other techniques convolution neural
network detects few more number of events in a small dataset and in future we will use larger dataset to
increase the number of events. Table 3 shows the results of detecting events. In a table for kicking the ball
event totally 2o events were considered, 13 out of 15 events were successfully detected but 3 events were
missed because when the player kick the ball it reaches another player less than 3 frames. For ball
possession event we have taken 10 events, 9 events were detected correctly. In missing cases the 2 players
are very close to one another. Event detection happen when the players were met in an appropriate number
of repeated frames. For instance, if the distance among the player and the ball is less than 3 frames we go for
event type as ball possession.

Table 3. Results for Detection of Event

Missed
Detected Events Total Accuracy
Detection

Ball Possession 10 9 92%

Kicking the ball 15 13 81%

Pass the ball 14 11 84%

Shot on goal 8 7 93%

While keeping this as a base we also detected the highly complex events like pass the ball and shot on the
goal. As a result, we merge the simple low-level events and high-level events using temporal and logical
operators. To classify the highly composite event, we have considered different cases at various abstraction
levels. Pass the ball events are detected when simple low-level events occur and it can use for referring to
the players of the same team. For example, consider player 1 and player 2 of the same team. The player 1
pass the ball to the player 2 as referring both players are same team this detection had been done. This can

ISSN: 2582-3981 [Link]

74
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
be achieved by the jersey color of the players. When we obtained a successive sequence of frame, event
detection starts, once the Player1 owns the ball is recognized it calculate the distance among the ball and the
player1 if it is low for a small number of frames after that distance among the ball and the player1 raises in
number of succeeding frames it checks for simple low-level Kicking the ball events. When the distance
between the ball and the player2 decreases it is highly complex “Pass the ball events”.

(23, 𝑃𝑎𝑠𝑠, 𝑡 + 𝑘̅ , (𝑝𝑎𝑠𝑠𝑖𝑛𝑔𝑃𝑙𝑎𝑦𝑒𝑟, 𝑝1 ), (𝑟𝑒𝑐𝑒𝑖𝑣𝑖𝑒𝑑𝑃𝑙𝑎𝑦𝑒𝑟, 𝑝2 ), (𝑝𝑎𝑠𝑠𝑒𝑑𝑂𝑏𝑗𝑒𝑐𝑡, 𝑏𝑎𝑙𝑙)) (7)

Where, 23 is the variable, Pass is the type of event, is duration of time in which the event exists.
Received player is the task execute by player p2 on the ball. Passing player is the task execute by player p1
on the ball.

In “shot on goal” event type consider three entities goal post, player and ball. The highly complex “Shot on
goal” occurs, when the player kick the ball, the distance among the player and the ball increases at the same
time the distance among the goal post and the ball loss the particular threshold.

𝐾𝑖𝑐𝑘𝑒𝑑𝑂𝑏𝑗𝑒𝑐𝑡 (8)
(20, 𝑆ℎ𝑜𝑡𝑂𝑛𝐺𝑜𝑎𝑙, +𝑘̅ , (𝐾𝑖𝑐𝑘𝑖𝑛𝑔𝑃𝑙𝑎𝑦𝑒𝑟, 𝑝), ( , 𝑏𝑎𝑙𝑙) , (𝐺𝑜𝑎𝑙𝑃𝑜𝑠𝑡, 𝐺))
𝑡 + 𝑘̅

Where, 20 is the variable, ”short on the goal” is the type of event, Kicking player is the task execute by
player P , GoalPost is the task of object G, while the ball motion towards goalpost.

6(a)

6(b)

Fig.6(a) & (b) Final Results

ISSN: 2582-3981 [Link]

75
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020

Fig.7 Confusion matrix

Table 4. Performance Measure

Parameters SVM RNN CNN

Accuracy 79.4% 84.5% 87.6%

Sensitivity 92.1% 81.7% 76.8%

Specificity 77.6% 80.7% 86.2%

7. Conclusion

There are numerous events present in the football games and the proposed work tends to classify the four
events of the football games. The implemented work attains the accuracy of 87.63% with the reduction in
the sensitivity and the specificity range of 72.5 and 86.2%. The confusion matrix of the proposed work
gives the classification accuracy during the training phase. For each mass category, the bounding boxes
related with a confidence score taken as input for event detection. The low-level complex events are
successfully detected by the system such as: "Ball possession" and "Kicking the ball ". The final results
indicate the effectiveness and validity of the proposed technique. The future work implies the consideration
of more number of events within the classification process.

Declarations

Source of Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or
not-for-profit sectors.

Competing Interests Statement

The authors declare no competing financial, professional and personal interests.

Consent for publication

We declare that we consented for the publication of this research work.

ISSN: 2582-3981 [Link]

76
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
Code availability

The programming code that we have used for this research is available and authors are willing to share
when it is required.

References
[1] Carcel Folch, E. (2011). Rich internet application for semi-automatic annotation of semantic shots on
keyframes.

[2] Hern, E., Beltr, Z. Z., & Zannatha, J. M. I. (2011, November). Soccer ball speed estimation using optical
flow for humanoid soccer player. In 2011 IEEE Electronics, Robotics and Automotive Mechanics
Conference (pp. 178-183). IEEE.

[3] Mackenzie, R., & Cushion, C. (2013). Performance analysis in professional soccer: Player and coach
perspectives. In Performance analysis of sport IX (pp. 49-57). Routledge.

[4] Kjellström, H., Kragić, D., & Black, M. J. (2010, June). Tracking people interacting with objects. In
2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (pp. 747-754).
IEEE.

[5] Netcraft, L. T. D. (2009). Netcraft web server survey. URL [Link]

[6] Priya, G. L., & Domnic, S. (2011, June). Video cut detection using Hilbert transform and GLCM. In
2011 International Conference on Recent Trends in Information Technology (ICRTIT) (pp. 749-753).
IEEE.

[7] de Sousa, S. F., Araújo, A. D. A., & Menotti, D. (2011, January). An overview of automatic event
detection in soccer matches. In 2011 IEEE Workshop on Applications of Computer Vision (WACV) (pp.
31-38). IEEE.

[8] Molchanova, V. S., Vidishcheva, E. V., & Stojsavljević, R. (2013). Designing Republic of Abkhazia
Tourism Site Web 2.0 Technology-Based as a Key Competitive Advantage of the Recreational Industry.
European Journal of Economic Studies, (4), 215-220.

[9] Zhu, G., Xu, C., Huang, Q., Rui, Y., Jiang, S., Gao, W., & Yao, H. (2008). Event tactic analysis based on
broadcast sports video. IEEE Transactions on Multimedia, 11(1), 49-67.

[10] Xu, P., Xie, L., Chang, S. F., Divakaran, A., Vetro, A., & Sun, H. (2001, August). Algorithms And
System For Segmentation And Structure Analysis In Soccer Video. In ICME (Vol. 1, pp. 928-931).

[11] Xie, L., Xu, P., Chang, S. F., Divakaran, A., & Sun, H. (2004). Structure analysis of soccer video with
domain knowledge and hidden Markov models. Pattern Recognition Letters, 25(7), 767-775.

[12] Wan, K., Yan, X., Yu, X., & Xu, C. (2003, November). Real-time goal-mouth detection in MPEG
soccer video. In Proceedings of the eleventh ACM international conference on Multimedia (pp. 311-314).

ISSN: 2582-3981 [Link]

77
 Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.4, Iss.3, Pages 66-78, July-September 2020
[13] Shimawaki, T., Sakiyama, T., Miura, J., & Shirai, Y. (2006, August). Estimation of ball route under
overlapping with players and lines in soccer video image sequence. In 18th International Conference on
Pattern Recognition (ICPR'06) (Vol. 1, pp. 359-362). IEEE.

[14] Nunez, J. R., Facon, J., & de Souza Brito, A. (2008, June). Soccer Video Segmentation: referee and
player detection. In 2008 15th International Conference on Systems, Signals and Image Processing (pp.
279-282). IEEE.

[15] Liang, D., Liu, Y., Huang, Q., & Gao, W. (2005, November). A scheme for ball detection and tracking
in broadcast soccer video. In Pacific-Rim Conference on Multimedia (pp. 864-875). Springer, Berlin,
Heidelberg.

[16] Leo, M., Mosca, N., Spagnolo, P., Mazzeo, P. L., & Distante, A. (2008, September). Real-time
multi-view event detection in soccer games. In 2008 Second ACM/IEEE International Conference on
Distributed Smart Cameras (pp. 1-10). IEEE.

[17] Cho, J. U., Jin, S. H., Dai Pham, X., Jeon, J. W., Byun, J. E., & Kang, H. (2006, October). A real-time
object tracking system using a particle filter. In 2006 IEEE/RSJ international conference on intelligent
robots and systems (pp. 2822-2827). IEEE.

[18] Kataoka, H., Hashimoto, K., & Aoki, Y. (2011, September). Player position estimation by monocular
camera for soccer video analysis. In SICE Annual Conference 2011 (pp. 1985-1990). IEEE.

[19] von Hoyningen-Huene, B. (2011). Real-time Tracking of Player Identities in Team Sports (Doctoral
dissertation, Technische Universität München).

[20] Hirano, S., & Tsumoto, S. (2004, September). Finding interesting pass patterns from soccer game
records. In European Conference on Principles of Data Mining and Knowledge Discovery (pp. 209-218).
Springer, Berlin, Heidelberg.

ISSN: 2582-3981 [Link]

78