A Feature-based Approach for Identifying Soccer Moves using an

Accelerometer Sensor

Omar Alobaid and Lakshmish Ramaswamy

Department of Computer Science, University of Georgia, Athens, Georgia, U.S.A.

Keywords:

Soccer, Activity Recognition, Accelerometer, Sensor.

Abstract:

During the past decade, Human Activity Recognition (HAR) systems have been an evolving topic due to the

popularity of smart devices. Recognizing soccer moves in real-time is an important research problem that has

not yet been studied thoroughly in the literature. In contrast to daily physical activities, recognizing soccer

moves poses a set of unique challenges, such as pattern irregularity and body positions when performing

these moves. In this paper, our goal is to recognize soccer moves in real-time by utilizing accelerometer

data. We explore three different feature-based algorithms: Time Series Forest, Fast Shapelets, and Bag-of-

SFA-Symbols. We also examine different factors that can affect the performance of these algorithms, such

as parameter tuning and accelerometer axis elimination. Additionally, we introduce a novel collaborative

model consisting of the above-mentioned algorithms in a majority voting mechanism to further enhance the

performance of the system. We also add a light-weight classiﬁer to act as a tie breaker in case of disagreement

between the classiﬁers. We experimentally choose the right parameters to reduce the training time drastically

without forfeiting the level of accuracy. Our collaborative model outperforms the single model by 2% to reach

84% in accuracy with a decrease in the training time by one order of magnitude.

1 INTRODUCTION

In the past decade, there has been rapid development

of Ubiquitous Computing involvement in our daily

lives. Smart phones, smart watches, and smart clothes

are examples of this technological explosion. These

smart devices are usually equipped with sensors that

can be utilized to serve numerous purposes. One of

these sensors is the Accelerometer, which measures

the change of velocity in m/s

in 3 dimensions. Be-

cause of its low power consumption, the accelerome-

ter sensor is frequently used to achieve different tasks,

such as device orientation and user’s physical activity.

Accelerometer data can be viewed as a time series

since it is a sequence of data points that are observed

over time. Time Series analysis is a major subject of

interest within the area of Data Mining. Time Series

analysis tasks include forecasting (De Gooijer and

Hyndman, 2006), querying (Hochheiser and Shnei-

derman, 2004), clustering (Liao, 2005), and classi-

ﬁcation (Xi et al., 2006). Time Series classiﬁca-

tion has been used by researchers in various domains,

such as in medical (Kurbalija et al., 2014), biologi-

cal (Tapinos, 2013), and geographical (Campbell and

Diebold, 2005) sciences.

In recent years, there has been rising interest in

recognizing human activity by using mobile sensors

(Anguita et al., 2012) (Bayat et al., 2014) (Wan et al.,

2015). Human Activity Recognition (HAR) systems

can be delineated into three categories (Lara et al.,

2013): external sensors, wearable sensors, and hy-

brid. Our focus in this paper is on wearable systems

for the cost, size, and convenience of these systems

compared to external sensor systems.

Soccer, as one of the most popular sports around

the world, also offers a unique opportunity for real-

time application software. There are many real-time

applications that track human activity, such as Apple

Health and Google Fit. However, none of these ap-

plications recognize soccer moves in real-time. Im-

plementing an affordable system that recognizes and

tracks basic soccer moves in real-time is desirable to

develop players’ soccer skills. As thousands of play-

ers practice soccer every day, having a real-time de-

tection system would assist in tracking this multitude

of training sessions. Our system can be extended to

analyze players’ performance. For example, coaches

may instruct players to perform a number of shots in

order to improve shooting skills. Instead of manually

counting the number of shots, those players can use

the proposed system to track number of shots taken in

real-time.

Alobaid, O. and Ramaswamy, L.

A Feature-based Approach for Identifying Soccer Moves using an Accelerometer Sensor.

DOI: 10.5220/0008910400340044

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 5: HEALTHINF, pages 34-44

ISBN: 978-989-758-398-8; ISSN: 2184-4305

Besides typical HAR system challenges, soccer

poses other unique challenges. First, soccer moves

have fast and irregular patterns, which make them

harder to distinguish compared to other daily physi-

cal activities. Second, players might use either foot

while playing, which makes the determination of sen-

sor placement a crucial decision. Our research goal is

to utilize accelerometer data to recognize basic soccer

moves performed by a player in a training session in

real-time without extra hardware.

In this paper, we applied feature-based algorithms

which typically extract meaningful features from the

data to be used in classiﬁcation. The contributions

of this paper include: ﬁrst, evaluate three differ-

ent feature-based algorithms: Time Series Forest,

Fast Shapelets, and Bag-of-SFA-Symbols, which rep-

resent different approaches- interval, Shapelet, and

Dictionary-based- to recognize soccer moves; second,

analyze the performance of these algorithms in terms

of accuracy and training time when the parameters are

tuned; third, we propose a voting approach composed

from the aforementioned algorithms to enhance accu-

racy. From this study, our results show that soccer

moves can be recognized in real-time with an accu-

racy of 84%.

2 RELATED WORK

Wearable sensor systems can be divided into two

types based on the learning approach: semi-

supervised and supervised.

• Semi-supervised: the model uses labeled and un-

labeled data in the training phase. This approach

is attractive to researchers because having labeled

data requires human resources to label it. Human

labeling becomes an unpractical approach when

the dataset is particularly large. However, obtain-

ing unlabeled data is easier, and it eliminates la-

beling cost (Guan et al., 2007). The authors in

(Radu et al., 2014) aimed to detect whether the

user is indoor or outdoor by employing a semi-

supervised approach as follows: use some of the

labeled data to assign a label(s) to clustered unla-

beled data; train a classiﬁer on small label data,

then tune the classiﬁer based on the unlabeled

data; and utilize collaborative learning, where

classiﬁers can enhance their performance by mu-

tual learning. In (Ghazvininejad et al., 2011), the

authors used a small portion of the labeled data in

a graph-based method. They calculated the asso-

ciation probability of each class using a k-nearest

neighbor graph. Then, these probabilities were

fed into the Hidden Markov Model to classify un-

seen examples.

• Supervised: the model uses labeled data only in

the training. This approach is the most popular

approach in HAR systems. In (Lee et al., 2017),

walking and running were identiﬁed with an ac-

curacy of 92% by training a Convolutional Neu-

ral Network. The researchers in (Kwapisz et al.,

2011) recognized daily activities by training logis-

tic regressions, decision trees, and multilayer neu-

ral network classiﬁers. Similarly, (Yazdansepas

et al., ), the researchers utilized a Shapelet-based

approach to recognize ambulatory activities in

real time with a high accuracy compared to off-

line systems.

3 MOTIVATION &

BACKGROUND

Nowadays, there are many commercial HAR systems

that help users to track their physical activity in real-

time, such as Fitbit, Apple, Garmin, and Android

watches. However, these watches are usually limited

to a number of general activities like running, walk-

ing, and swimming. Our goal in this research is to

recognize a different and more intense type of sport

in real-time. For this, we chose soccer, because it is

the world’s most popular sport (Dunning, 2013). With

millions around the world playing this sport, an af-

fordable system, that can track players’ soccer moves

during training sessions in real time, can help to im-

prove players’ performances. Furthermore, to the best

of our knowledge, identifying soccer actions in real

time using the accelerometer sensor has not yet been

discussed in the literature. It is worth mentioning

that soccer moves are harder to recognize compared

to daily activities, because soccer moves are fast and

irregular, compared to activities like walking and run-

ning. A related point to consider is that players have

different techniques to perform these moves. Build-

ing a player-independent platform to recognize these

moves is a signiﬁcant feature of our system.

Our research objective is to build a player-

independent platform to identify soccer moves in

real-time utilizing accelerometer data. One of the

possible approaches for real-time classiﬁcation is to

apply time series classiﬁcation algorithms. We apply

lightweight feature-based algorithms to classify

streaming data on-the-ﬂy with minimal overhead

on resource-constrained mobile devices. Though

machine learning algorithms are most popular in

HAR studies, most of these algorithms require high

A Feature-based Approach for Identifying Soccer Moves using an Accelerometer Sensor

computational resources to extract and determine

the most important features. As such, feature-based

algorithms are more efﬁcient when paired with

accelerometer data on mobile devices.

In our research, we focused on the following moves:

shooting the ball, passing the ball, heading the

ball, running, and dribbling. We chose these moves

because these are the basic moves in soccer.

Time series classiﬁcation is a classic problem in

the Time Series analysis domain. Classifying un-

known instances is a desired goal in real-life appli-

cations. There are two main approaches to classify

time series (Baydogan et al., 2013): instance-based

and feature-based. In this section we will discuss the

instance-based approach. The feature-based approach

will be discussed in the next section.

3.1 Instance-based

This approach uses a similarity measure between the

new instance and the training instances. Euclidean

Distance and Dynamic Time Warping are examples

of this approach. The instance-based approach is

characterized by its simplicity. On the other hand,

it only works well with short time series and does

not generalize well to long and noisy time series

(Sch

afer, 2015).

3.1.1 Euclidean Distance

This method measures the straight distance between

two data points in the Euclidean space.

In 3-dimensional space, calculating the Euclidean

Distance between two points from the accelerometer

sensor is shown in Equation 1:

ED(p,q) =

−q

)

+ (p

−q

)

+ (p

−q

)

(1)

Euclidean Distance is widely used due to its efﬁ-

ciency and simplicity. However, it does not perform

well in most cases due its sensitivity to distortion

(Ratanamahatana and Keogh, 2004) and intolerance

to time shifting.

3.1.2 Dynamic Time Warping (DTW)

Dynamic Time Warping (Ratanamahatana and

Keogh, 2004) performs non-linear matching between

two series by reducing the distance between the time

series. This approach was proposed to overcome

Euclidean Distance’s weakness to distortion and time

shifting. Dynamic Time Warping showed superb

performance in many applications. Nonetheless, the

main disadvantage of Dynamic Time Warping is its

time complexity O(n

) where n is the length of the

time series. There are variations of DTW, such as

Weighted DTW (Jeong et al., 2011) and Time warp

edit distance (TWED) (Marteau, 2009). Weighted

DTW assigns less weight to points that have the

largest difference between a training point and a test-

ing point. This step aims to reduce the outliers effect

on the classiﬁcation. On the other hand, Time warp

edit distance (TWED) introduces a parameter called

stiffness which controls the ﬂexibility of TWED. The

Stiffness parameter compromises between inﬁnite

stiffness in Euclidean Distance and zero stiffness in

Dynamic Time Warping (DTW).

4 FEATURE-BASED SOCCER

MOVES IDENTIFICATION

The feature-based approach generates features from

time series to be compared instead of similarity mea-

sures in the instance-based method. Feature-based

is usually faster than instance-based when it uses

fast feature extraction and classiﬁcation algorithms

(Baydogan et al., 2013). Feature-based can be

divided into three types: Interval, Shapelets, and

Dictionary. The following taxonomy and examples

are adapted from (Bagnall et al., 2017).

4.0.1 Time Series Forest (TSF)

TSF (Deng et al., 2013) is a tree-ensemble that de-

ploys Random Forest sampling approach to reduce

the feature space for intervals. TSF samples

√

m in-

tervals where m = T S.length. Using this approach

reduces the feature space drastically from O(m

) to

O(m). For each interval, mean, standard deviation,

and slope are calculated. These features are used to

distinguish between different soccer moves. For ex-

ample, mean can be used to distinguish between head-

ing and passing the ball, since mean value of the ver-

tical axis in heading will be higher than passing (i.e.

most players jump to head the ball which results in

higher values on the vertical axis). TSF uses En-

tropy and Distance to spot best splits in the trees if the

node exceeds a threshold. The goal of the Entropy is

to determine the most discriminated/ disparate nodes

which can distinctly separate the classes. In many

cases, there will be more than one possible split. To

break this tie, the distance will be measured between

the candidate threshold and the nearest feature value.

HEALTHINF 2020 - 13th International Conference on Health Informatics

One of the main advantages of TSF is the ability to

train trees independently, which allows for parallel

training. To classify an unseen example, TSF assigns

a label based on a majority voting approach.

4.0.2 Fast Shapelets

Fast Shapelet (Rakthanmanon and Keogh, 2013) is a

heuristic algorithm. It converts the time series’ real

values into an alphabetical representation. The pur-

pose of the conversion is to reduce the search cost,

since the values range is limited. Another major

purpose of the conversion is to hash the data to in-

crease the search accuracy by utilizing collision his-

tory. In soccer, different moves have different signal

amplitudes which will result in different SAX repre-

sentations. For instance, passing, and shooting have

roughly similar signal patterns, but in different ampli-

tudes. Utilizing the SAX representation in this case

helps to differentiate between the two classes. In ad-

dition, the SAX representation helps reduce the vari-

ety of move patterns from each player.

Fast Shapelet uses a sliding window technique to con-

vert the data to a SAX representation. Because the

window size has a length less than the time series

segment, there will be multiple SAX words that re-

fer to the same time series segment. However, this

will lead to false dismissals where two nearly identi-

cal segments could create two different SAX words.

Random Projection was proposed by the authors to

solve this issue. For example, if there are words on

a length of 5, then Random Projection masks 2 po-

sitions to produce words on a length of 3. This ap-

proach will increase the chance that two similar SAX

words are mapped to the same masked word. After

r iterations, distinguishing power will be calculated

to differentiate between words that strongly represent

the same class while hardly appearing in any other

class.

4.0.3 Bag-of-SFA-Symbols

Bag-of-SFA-Symbols (BOSS) (Sch

afer, 2015) is a

bag-of-words model that uses a structural-based ap-

proach to extract representative features. BOSS con-

verts the raw time series into various substructures

by applying Symbolic Fourier Approximation (SFA).

SFA utilizes Fourier Transform and Multiple Coefﬁ-

cient Binning to achieve approximation and quanti-

zation. SFA reduces the noise of these substructures

by applying low pass ﬁltering to facilitate the work of

string matching algorithms.

BOSS applies a sliding window. Each window is

normalized for amplitude invariance, and the mean

normalization is enabled depending on the time series

nature. In soccer, the mean is an important feature to

separate between different moves. For instance, body

movement in shooting is more intense than passing,

which will increase the mean sensor reading values.

In Section 6, we showed the importance of the mean

on classiﬁcation accuracy. For each window, SFA

is applied. Because adjacent windows have a high

chance of being identical windows, numerosity reduc-

tion is applied to avoid over-counting a substructure.

Finally, BOSS creates a histogram of the SFA words.

A modiﬁed version of Euclidean Distance is used to

measure the distance between words. To classify any

new unseen instance, BOSS deploys a 1-NN search

algorithm to ﬁnd the nearest hit.

5 ARCHITECTURE

In this paper, we conduct comprehensive experi-

ments to recognize soccer movements and compare

the performance between 3 different feature-based al-

gorithms: Time series forest (TSF), Fast Shapelets

(FS) from, and Bag-of-SFA-Symbols (BOSS). The

main objective is to ﬁnd the most accurate, yet efﬁ-

cient algorithm to achieve the recognition tasks. Af-

ter we examine different algorithms, we propose a

novel collaborative model composed from the above-

mentioned algorithms. The following subsections ex-

plain the workﬂow of our system.

5.1 Data Collection

The data was collected from 16 players between 18

and 35 years old. Each player performed different

soccer actions: shooting the ball, passing the ball,

heading the ball, running, and dribbling. To collect

the data, a Samsung Galaxy S5 was used to record

the accelerometer data using the Sensor Kinetics Ap-

plication (INNOVENTIONS

, 2017) at a 100 Hz

sampling rate. Every player wore a belt to hold the

phone in the abdomen area. The central abdomen area

was chosen intentionally to help the system recognize

moves from players regardless of their dominant foot.

5.2 Signal Denoising

The accelerometer sensor is sensitive to even the

weakest movement, which leads to unwanted data in

recognition tasks. To overcome this issue, signal de-

noising/smoothing is applied. There are many sig-

nal denoising methods (Lorenz, ), such as Wavelt and

A Feature-based Approach for Identifying Soccer Moves using an Accelerometer Sensor

Moving Average Filter

Raw Accelerometer Data

3-Second Windows

Time

Series

Forest

Fast

Shapelets

Bag-of-

SFA-

Symbols

Decision

Tree

Yes

Distinct

Classes?

Select

Majority

Class

Time

Domain

Features

Figure 1: Architecture of the the proposed system.

Linear Fourier. Because our system is designed to

run on portable devices with limited computational

resources, we chose Moving Average Filter method

(Lorenz, ) as our smoothing method.

5.3 Signal Segmentation

Soccer practice sessions typically last between several

minutes to one hour or more. As a result, dividing the

incoming data into segments/windows is necessary to

extract the important features. We conducted several

preliminary experiments to ﬁnd the best window size

for this purpose. We found a 3-second window is the

optimal size for our system to capture the characteris-

tics of each soccer move without negatively affecting

the accuracy.

5.4 Classiﬁcation

Our goal is to examine 3 feature-based algorithms

that use different approaches. We explored Time

series forest (TSF) from the Intervals family, Fast

Shapelets (FS) from the Shapelets family, and Bag-

of-SFA-Symbols (BOSS) from the Dictionary family.

The resulting segments from the signal segmentation

step were then used to train the 3 classiﬁers. After

tuning each algorithm to ﬁnd the best parameters, we

proposed a novel approach to combine the aforemen-

tioned algorithms in voting mechanism to improve the

accuracy and reduce the training time. The main goals

of this step are to enhance the classiﬁcation accuracy

and reduce the over-ﬁtting probability. Our collabo-

rative model is independent, which allows for paral-

lel training. We also added a light-weight classiﬁer,

which will act as a tie breaker (e.g. when each clas-

siﬁer produces a distinct class). Decision Tree was

chosen as the tie breaker classiﬁer due to its speed and

efﬁciency. To train the fourth classiﬁer, we extracted

10 time domain features from the x and y axes.

6 EMPIRICAL EVALUATION

In all experiments, we used the open-source code

from the Time Series Classiﬁcation repository (Bag-

nall et al., 2017).

All of the experiments were done using a single

machine with a dual-core processor (4 logical proces-

sors) and 8 Gigabytes memory. To evaluate the clas-

siﬁcation performance, we used 10-fold cross valida-

tion.

6.1 Time Series Forest (TSF)

In our preliminary experiment, we used 500 trees and

30 candidates. The accuracy of our system was is

81%. Figure 2 shows accuracy per activity. It can be

clearly seen that shooting and passing have the high-

est accuracy compared to the rest. On the other hand,

heading has the lowest accuracy. We believe the rea-

son for these results is the nature of these activities.

Passing and shooting have distinct natures. When a

player passes the ball, the power applied on the ball is

considerably less than shooting the ball, which makes

body movements in passing less intense than shoot-

ing. Table 1 illustrates the confusion matrix.

Table 1: Confusion matrix of TSF.

body movements in passing less intense than shoot-

ing. Table 1 illustrates the confusion matrix.























    



Figure 2: TSF Accuracy per activity

Table 1: Confusion matrix of TSF

Running Passing Heading Shooting Dribbling

Running 0.80 0.04 0.06 0.00 0.10

Passing 0.00 0.88 0.02 0.10 0.00

Heading 0.08 0.17 0.70 0.06 0.00

Shooting 0.00 0.08 0.02 0.91 0.00

Dribbling 0.16 0.00 0.09 0.02 0.73

6.1.1 Parameters Effect

TSF has two parameters to be entered by the user:

number of trees and number of candidate thresh-

olds. In this experiment, we attempted to ﬁnd the

best parameters combination to achieve best accuracy

while lowering training time using a greedy approach

(Frank Hutter, )(Matuszyk et al., ). We started with

a varying number of trees while keeping the number

of candidates ﬁxed to 30. The forests’ sizes were

50, 100, 250, 500(De f ault), 1000. The results show

that the execution time increases as the number of

trees increases, while the improvement in accuracy is

less than 1% which is insigniﬁcant. Figure 3 shows

the detailed performance.































     











Figure 3: TSF performance with different number of trees

After testing the impact of the number of trees, we

evaluated the effect of the number of candidate thresh-

olds by ﬁxing the number of trees to 100 and 250,

as these numbers showed the highest accuracy while

carrying less training time. The tested number of can-

didate thresholds were 1, 3, 5, 10, 20(De f ault), 30, 50.

The results in Figure 4 show that increasing the num-

ber of candidates from 1 to 30 increases the accuracy

by ∼4%, from 77% to 81% in the case of 100 trees,

with a minimal rise in the execution time by ∼500 ms.

However, there was is no difference in accuracy when

the forest size is 250. Figure 4 shows the performance

of our experiments.

6.1.2 Accelerometer Axis Elimination Effect

One of our primary goals in this paper is to increase

the efﬁciency of our system while maintaining the

same level of accuracy. For this step, we tested re-

moving one axis at a time from the accelerometer data

HEALTHINF 2020 - 13th International Conference on Health Informatics

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

Running Passing Heading Shooting Dribbling

Accuracy

Figure 2: TSF Accuracy per activity.

6.1.1 Parameters Effect

TSF has two parameters to be entered by the user:

number of trees and number of candidate thresh-

olds. In this experiment, we attempted to ﬁnd the

best parameters combination to achieve best accuracy

while lowering training time using a greedy approach

(Frank Hutter, )(Matuszyk et al., ). We started with

a varying number of trees while keeping the number

of candidates ﬁxed to 30. The forests’ sizes were

50,100,250,500(De f ault),1000. The results show

that the execution time increases as the number of

trees increases, while the improvement in accuracy is

less than 1% which is insigniﬁcant. Figure 3 shows

the detailed performance.

5000

10000

15000

20000

25000

30000

35000

40000

0.6

0.65

0.7

0.75

0.8

0.85

25 50 100 250 500 1000

Training time/fold in ms

Accuracy

Number of trees

Accuracy

Training Time

Figure 3: TSF performance with different number of trees.

After testing the impact of the number of trees, we

evaluated the effect of the number of candidate thresh-

olds by ﬁxing the number of trees to 100 and 250,

as these numbers showed the highest accuracy while

carrying less training time. The tested number of can-

didate thresholds were 1,3,5,10,20(De f ault),30,50.

The results in Figure 4 show that increasing the num-

ber of candidates from 1 to 30 increases the accuracy

by ∼4%, from 77% to 81% in the case of 100 trees,

with a minimal rise in the execution time by ∼500 ms.

However, there was is no difference in accuracy when

the forest size is 250. Figure 4 shows the performance

of our experiments.

1000

2000

3000

4000

5000

6000

7000

8000

100 250

Training time/fold in ms

Number of trees

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

100 250

Accuracy

Number of trees

Figure 4: TSF performance with different number of candi-

date thresholds.

6.1.2 Accelerometer Axis Elimination Effect

One of our primary goals in this paper is to increase

the efﬁciency of our system while maintaining the

same level of accuracy. For this step, we tested re-

moving one axis at a time from the accelerometer data

in the forests with 100 and 250 trees while candidate

thresholds were set to 30. In our ﬁrst trial, the forest

size at 100 has an accuracy of 81% with all axes. Af-

ter eliminating the z axis, the accuracy decreases by

only 1%, while the training time drops by 45% from

∼4800 to ∼2700 ms. When we remove the x and y

axes, the accuracy sharply decreases to 75% and 73%,

respectively.

In our second trial, the forest size was 250. Remov-

ing the z axis does not affect the accuracy, while the

training time reduces by 23%, which is a signiﬁcant

improvement toward an efﬁcient recognition system.

However, using y,z and x,z axes affects the perfor-

mance negatively by 4% and 5%, respectively. Figure

5 shows the performance of our experiments.

6.1.3 Parallel Training Effect

Constructing and training a large forest sequentially

is not the optimal approach. Therefore, in this ex-

periment, we converted the TSF implementation into

a parallel approach in order to utilize the computation

resources. When forest size is small (i.e. 50 trees), the

A Feature-based Approach for Identifying Soccer Moves using an Accelerometer Sensor

2000

4000

6000

8000

10000

12000

100 250

Training time/fold in ms

Number of trees

all

x,y

x,z

y,z

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

100 250

Accuracy

Number of trees

all

x,y

x,z

y,z

Figure 5: TSF Accuracy and training time based on the used

axes.

training time is similar between the two approaches.

An improvement of 60% is achieved when the forest

size is larger than 250 trees. Figure 6 shows the per-

formance of our experiments.

5000

10000

15000

20000

25000

30000

50 100 250 500 750 1000

Training time/fold in ms

Number of Trees

Parallel

Sequential

Figure 6: TSF performance with parallel training.

6.2 Fast Shapelets

In our exploratory experiments, we used the default

parameters as shown in the next paragraph. The ac-

curacy is 74%, but the training time is substantially

longer (i.e. 110 mins/fold) compared to TSF. It is

worth mentioning that our data is relatively small

compared to other datasets, meaning that training

time will increase dramatically with larger datasets.

The reason behind this slowness is FS does an in-

tensive search to construct Shapelet from all possi-

ble lengths. In the next section, we discuss the effect

of reducing the possible Shapelet lengths. Figure 7

shows accuracy per activity, and Table 2 shows the

confusion matrix.

0.000

0.100

0.200

0.300

0.400

0.500

0.600

0.700

0.800

0.900

1.000

Running Passing Heading Shooting Dribbling

Accuracy

Figure 7: FS Accuracy per activity.

Table 2: Confusion matrix of FS.















 



































 















Figure 5: TSF Accuracy and training time based on the used

axes

training time is similar between the two approaches.

An improvement of 60% is achieved when the forest

size is larger than 250 trees. Figure 6 shows the per-

formance of our experiments.















     









Figure 6: TSF performance with parallel training

6.2 Fast Shapelets

In our exploratory experiments, we used the default

parameters as shown in the next paragraph. The ac-

curacy is 74%, but the training time is substantially

longer (i.e. 110 mins/fold) compared to TSF. It is

worth mentioning that our data is relatively small

compared to other datasets, meaning that training

time will increase dramatically with larger datasets.

The reason behind this slowness is FS does an in-

tensive search to construct Shapelet from all possi-

ble lengths. In the next section, we discuss the effect

of reducing the possible Shapelet lengths. Figure 7

shows accuracy per activity, and Table 2 shows the

confusion matrix.























    



Figure 7: FS Accuracy per activity

Table 2: Confusion matrix of FS

Running Passing Heading Shooting Dribbling

Running 0.55 0.00 0.10 0.00 0.35

Passing 0.00 0.94 0.03 0.02 0.02

Heading 0.06 0.09 0.64 0.21 0.00

Shooting 0.02 0.04 0.10 0.82 0.02

Dribbling 0.25 0.00 0.00 0.02 0.73

6.2.1 Parameters Effect

Fast Shapelet has many default parameters that can

be modiﬁed: r = 10, which is the number of itera-

tions to perform random masking, top

k = 10, which

is the top k

subsequences that have the highest score,

min haplet len = 10, which is the minimum length of

any Shapelet, and step = 1, which is the increment in

the Shapelet discovery search.

Our initial experiments showed that step parameter is

the most inﬂuential factor on training time. As a re-

sult, we started by testing 10 various values for step

in the interval of [1, 500] while the other parameters

were left unchanged. Our experiments showed that

the accuracy is not drastically affected by increasing

the step in most cases, while the improvement in the

training time is exponential. For example, when we

increased the step from 1 to 5, the accuracy was low-

ered by only 1%, while the training time was reduced

by ∼80%. When step = 30 or 50, the accuracy is

74% and 75%, respectively, while the time is reduced

by one order of magnitude. Figure 8 shows the impact

of step size on the accuracy and training time.

Next, we set the value of step to 50, since

this value showed the best performance. We

6.2.1 Parameters Effect

Fast Shapelet has many default parameters that can

be modiﬁed: r = 10, which is the number of itera-

tions to perform random masking, top k = 10, which

is the top k

subsequences that have the highest score,

min shaplet len = 10, which is the minimum length

of any Shapelet, and step = 1, which is the increment

in the Shapelet discovery search.

Our initial experiments showed that step parameter is

the most inﬂuential factor on training time. As a re-

sult, we started by testing 10 various values for step

in the interval of [1,500] while the other parameters

were left unchanged. Our experiments showed that

the accuracy is not drastically affected by increasing

the step in most cases, while the improvement in the

training time is exponential. For example, when we

increased the step from 1 to 5, the accuracy was low-

ered by only 1%, while the training time was reduced

by ∼80%. When step = 30 or 50, the accuracy is

74% and 75%, respectively, while the time is reduced

by one order of magnitude. Figure 8 shows the impact

of step size on the accuracy and training time.

Next, we set the value of step to 50, since

this value showed the best performance. We

then tested 6 values for top k which were

{1,3,5,10(De f ault),20,30}. Our results show

that the default value, 10, is the highest value in

HEALTHINF 2020 - 13th International Conference on Health Informatics

100

1000

10000

100000

1000000

10000000

0.5

0.55

0.6

0.65

0.7

0.75

0.8

1 5 10 30 50 75 100 150 225

Training Time/fold in ms (log scale)

Accuracy

Step

Accuracy

Training Time

Figure 8: FS Accuracy and training time based on step size.

accuracy. Increasing the values to 20 or 30 affect

the accuracy and training time negatively in both

step values. On the other hand, reducing the value

of top k to 3 reduced the training time by 50% with

only a 2% accuracy loss, which is suitable for limited

computational resources. Figure 9 illustrates our

results in regards to the top k values.

50000

100000

150000

200000

250000

300000

350000

400000

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

1 3 5 10 20 30

Training Time/fold in ms

Accuracy

top_k

Accuracy

Training Time

Figure 9: FS Accuracy based on step size (50) and top k

values.

Random Projection iteration r is another parame-

ter that we can modify. We tested 3 values for r, which

were {5, 10(De f ault),20}. Our results show that r

does not affect the accuracy, but when r increased,

the training time also increases by 10%. Figure 10

illustrates the results of this experiment.

6.2.2 Accelerometer Axis Elimination Effect

In this section, we examined the impact of an axis

elimination while ﬁxing the parameters as follows:

step = 50, r = 5, and top k = 3,5,10. When only

x,y is used, the accuracy drops by 3%, from 73% to

70%, with a 66% reduction in the training time. In

the case of using x,z and y,z, the accuracy drops by

10% and 30%, respectively. These results conﬁrmed

what we found in the TSF experiment about the im-

portance of the x axis in the recognition task. Using

only two axes, x,y, can lead to an acceptable accuracy

in order to shorten the training time. Figure 11 shows

the results of using different top k values.

20000

40000

60000

80000

100000

120000

140000

160000

3 5 10

Training Time/fold in ms

top_k

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

3 5 10

Accuracy

top_k

Figure 10: FS Accuracy and training time based on step size

(50) and top k values (3,5,10) with different r values.

6.3 Bag-of-SFA-Symbols

In our ﬁrst BOSS experiment, we used the default pa-

rameters to train our model. The accuracy is 82%,

which is similar to TSF, however, the training time is

considerably longer than TSF (i.e. 52 mins/fold). Fig-

ure 12 shows accuracy per activity, and Table 3 shows

the confusion matrix.

Table 3: Confusion matrix of BOSS.































        











Figure 8: FS Accuracy and training time based on step size

then tested 6 values for top

k which were

{1, 3, 5, 10(De f ault), 20, 30}. Our results show

that the default value, 10, is the highest value in

accuracy. Increasing the values to 20 or 30 affect

the accuracy and training time negatively in both

step values. On the other hand, reducing the value

of top k to 3 reduced the training time by 50% with

only a 2% accuracy loss, which is suitable for limited

computational resources. Figure 9 illustrates our

results in regards to the top k values.





































     













Figure 9: FS Accuracy based on step size (50) and top k

values

Random Projection iteration r is another parame-

ter that we can modify. We tested 3 values for r, which

were {5, 10(De f ault), 20}. Our results show that r

does not affect the accuracy, but when r increased,

the training time also increases by 10%. Figure 10

illustrates the results of this experiment.

6.2.2 Accelerometer Axis Elimination Effect

In this section, we examined the impact of an axis

elimination while ﬁxing the parameters as follows:

step = 50, r = 5, and top

k = 3, 5, 10. When only

x, y is used, the accuracy drops by 3%, from 73% to

70%, with a 66% reduction in the training time. In

the case of using x, z and y, z, the accuracy drops by

10% and 30%, respectively. These results conﬁrmed

what we found in the TSF experiment about the im-

portance of the x axis in the recognition task. Using

only two axes, x, y, can lead to an acceptable accuracy



















  































  













Figure 10: FS Accuracy and training time based on step size

(50) and top

k values (3,5,10) with different r values

in order to shorten the training time. Figure 11 shows

the results of using different top

k values.

6.3 Bag-of-SFA-Symbols

In our ﬁrst BOSS experiment, we used the default pa-

rameters to train our model. The accuracy is 82%,

which is similar to TSF, however, the training time is

considerably longer than TSF (i.e. 52 mins/fold). Fig-

ure 12 shows accuracy per activity, and Table 3 shows

the confusion matrix.

Table 3: Confusion matrix of BOSS

Running Passing Heading Shooting Dribbling

Running 0.87 0.00 0.00 0.04 0.09

Passing 0.00 0.89 0.03 0.08 0.00

Heading 0.00 0.17 0.82 0.02 0.00

Shooting 0.00 0.12 0.08 0.79 0.01

Dribbling 0.14 0.00 0.07 0.00 0.80

6.3.1 Parameters Effect

BOSS has few parameters that can be modiﬁed.

al pha

size = 4 refers to the number of letters in the

string representation. mean norm is a boolean vari-

able to determine whether to perform mean normal-

ization or not. window

length is a dynamic vari-

able that can be affected by 3 different parameters:

min

window length = 10, max window length =

6.3.1 Parameters Effect

BOSS has few parameters that can be modiﬁed.

al pha size = 4 refers to the number of letters in the

string representation. mean norm is a boolean vari-

able to determine whether to perform mean normal-

ization or not. window length is a dynamic vari-

able that can be affected by 3 different parameters:

min window length = 10, max window length =

T S.length, and step = 1, which is the increase in

the window’s size. The default parameters will con-

duct SFA transforming to all possible window sizes

1,2,3,...,T S.length, which is an expensive computa-

tion. In this experiment, we used different step values

A Feature-based Approach for Identifying Soccer Moves using an Accelerometer Sensor

20000

40000

60000

80000

100000

120000

140000

3 5 10

Training Time/fold in ms

top_k

All

X,Y

X,Z

Y,Z

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

3 5 10

Accuracy

top_k

All

X,Y

X,Z

Y,Z

Figure 11: FS Accuracy and training time based on the used

axes.

0.00

0.10

0.20

0.30

0.40

0.50

0.60

0.70

0.80

0.90

1.00

Running Passing Heading Shooting Dribbling

Accuracy

Figure 12: BOSS Accuracy per activity.

to measure the classiﬁcation performance in terms of

accuracy and training time. Our results show that in-

creasing the step value to 15 does not affect the accu-

racy by more than 2%, however it reduces the training

time by one order of magnitude (i.e. from 52 mins to 4

mins/fold). From Figure 13, we can conclude that any

step 6 25 can achieve an acceptable accuracy, while

signiﬁcantly reducing the training time.

The next experiment examined the effect of

al pha size. We tested the following values

3,4(De f ault),6,8,10 when step = 5 and step = 15.

Our experiments show that 4 is the best alphabet size

and increasing the alphabet size leads to a sharp de-

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 5 10 15 25 50 75 100

100

1000

10000

100000

1000000

Accuracy

Step

Training Time/fold in ms

(log scale)

Trainig Time

Accuracy

Figure 13: BOSS training time and accuracy based on step

size.

crease in accuracy. When we increase the al pha size

from 4 to 6, the accuracy drops by 7% when step = 5,

and by 11% when step = 15. These results are con-

sistent with the ﬁndings in (Lin et al., 2012) (Sch

afer,

2015), which suggest 4 is an appropriate alphabet

size.

Our last experiment in parameter tuning was the

mean normalization effect. For every SFA word, the

ﬁrst Fourier coefﬁcient was not included. Our results

show that the normalization crucially lowered the ac-

curacy by more than 11%. Our interpretation for this

result is that soccer movements are fast and have high

amplitude, and the normalization reduces the signal

amplitude which leads to confusion in the classiﬁca-

tion process. Figure 14 summarizes the results of the

last two experiments.

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

3 4 6 8 10 3 4 6 8 10

5 15

Accuracy

Alphapets size

step size

YES

Figure 14: BOSS Accuracy based on the alphabet size (3,

4, 6, 8, 10) and mean normalization (yes,no).

6.3.2 Accelerometer Axis Elimination Effect

In this experiment, we tested the effect of elimi-

nating one of the three axes. We used the fol-

lowing attributes: al pha size = 4, mean

orm =

f alse, step = 15, min window length = 10, and

max window length = T S.length. Our results show

that eliminating the z axis causes a decrease of 3%,

from 82% to 79%, while reducing the training time by

more than 50%. Eliminating the y and x axes reduces

the accuracy by 8% and 11%, respectively. These re-

HEALTHINF 2020 - 13th International Conference on Health Informatics

sults align with the outcomes of TSF and FS exper-

iments, which reveals the importance of x,y axes in

the classiﬁcation process, compared with the z axis,

which can be omitted without crucially affecting the

accuracy. Figure 15 shows the accuracy and training

time of the axis elimination experiment.

100000

200000

300000

400000

500000

600000

700000

5 15

Training time/fold in ms

Step

all

x,y

x,z

y,z

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

5 15

Training time/fold in ms

Step

all

x,y

x,z

y,z

Figure 15: BOSS accuracy and training time based on the

used axes.

6.4 Collaborative Feature-based

Approach

Our collaborative model achieves better accuracy by

2% with less training overhead, because we selected

the best possible parameters for these models. Fig-

ure 16 shows the accuracy of the collaborative model

compared to TSF, FS, and BOSS classiﬁers.

Accuracy

TSF FS BOSS Proposed System

Figure 16: Collaborative Approach compared to single

models.

7 CONCLUSION

In this paper, we aimed to recognize soccer

moves in real-time. We comprehensively ex-

amined three different feature-based approaches,

which are Time Series Forest (interval-based),

Fast Shapelets (Shapelet-based), and Bag-of-SFA-

Symbols (Dictionary-based). We studied different

factors that might affect the accuracy and the train-

ing time, such as parameters tuning and axis elimi-

nation. We tuned our model to reduce the training

time by one order of magnitude, in the case of Fast

Shapelets and Bag-of-SFA-Symbols, without sacriﬁc-

ing the accuracy. We then proposed a collaborative

model where we combined all three approaches in a

voting mechanism using only two axes, which led to

an increase in accuracy by 2% to reach 84%.

REFERENCES

Anguita, D., Ghio, A., Oneto, L., Parra, X., and Reyes-

Ortiz, J. L. (2012). Human activity recognition

on smartphones using a multiclass hardware-friendly

support vector machine. In International workshop on

ambient assisted living, pages 216–223. Springer.

Bagnall, A., Lines, J., Bostrom, A., Large, J., and Keogh,

E. (2017). The great time series classiﬁcation bake

off: a review and experimental evaluation of recent

algorithmic advances. Data Mining and Knowledge

Discovery, 31(3):606–660.

Bayat, A., Pomplun, M., and Tran, D. A. (2014). A study

on human activity recognition using accelerometer

data from smartphones. Procedia Computer Science,

34:450–457.

Baydogan, M. G., Runger, G., and Tuv, E. (2013). A bag-

of-features framework to classify time series. IEEE

transactions on pattern analysis and machine intelli-

gence, 35(11):2796–2802.

Campbell, S. D. and Diebold, F. X. (2005). Weather fore-

casting for weather derivatives. Journal of the Ameri-

can Statistical Association, 100(469):6–16.

De Gooijer, J. G. and Hyndman, R. J. (2006). 25 years of

time series forecasting. International journal of fore-

casting, 22(3):443–473.

Deng, H., Runger, G., Tuv, E., and Vladimir, M. (2013). A

time series forest for classiﬁcation and feature extrac-

tion. Information Sciences, 239:142–153.

Dunning, E. (2013). Sport matters: Sociological studies of

sport, violence and civilisation. Routledge.

Frank Hutter, M. L. Algorithm conﬁguration:

Hands-on tutorial. http://www.ml4aad.org/wp-

content/uploads/2016/02/AC-Tutorial.pdf.

Ghazvininejad, M., Rabiee, H. R., Pourdamghani, N., and

Khanipour, P. (2011). Hmm based semi-supervised

learning for activity recognition. In Proceedings of

the 2011 international workshop on Situation activity

& goal awareness, pages 95–100. ACM.

A Feature-based Approach for Identifying Soccer Moves using an Accelerometer Sensor

Guan, D., Yuan, W., Lee, Y.-K., Gavrilov, A., and Lee, S.

(2007). In Embedded and Real-Time Computing Sys-

tems and Applications, 2007. RTCSA 2007. 13th IEEE

International Conference on, pages 469–475. IEEE.

Hochheiser, H. and Shneiderman, B. (2004). Dynamic

query tools for time series data sets: timebox widgets

for interactive exploration. Information Visualization,

3(1):1–18.

INNOVENTIONS

, I. ((accessed February 3, 2017)). Sen-

sor Kinetics Pro.

Jeong, Y.-S., Jeong, M. K., and Omitaomu, O. A. (2011).

Weighted dynamic time warping for time series clas-

siﬁcation. Pattern Recognition, 44(9):2231–2240.

Kurbalija, V., Radovanovi

c, M., Ivanovi

c, M., Schmidt, D.,

von Trzebiatowski, G. L., Burkhard, H.-D., and Hin-

richs, C. (2014). Time-series analysis in the medical

domain: A study of tacrolimus administration and in-

ﬂuence on kidney graft function. Computers in biol-

ogy and medicine, 50:19–31.

Kwapisz, J. R., Weiss, G. M., and Moore, S. A. (2011).

Activity recognition using cell phone accelerometers.

ACM SigKDD Explorations Newsletter, 12(2):74–82.

Lara, O. D., Labrador, M. A., et al. (2013). A survey

on human activity recognition using wearable sen-

sors. IEEE Communications Surveys and Tutorials,

15(3):1192–1209.

Lee, S.-M., Yoon, S. M., and Cho, H. (2017). Human activ-

ity recognition from accelerometer data using convo-

lutional neural network. In Big Data and Smart Com-

puting (BigComp), 2017 IEEE International Confer-

ence on, pages 131–134. IEEE.

Liao, T. W. (2005). Clustering of time series data—a survey.

Pattern recognition, 38(11):1857–1874.

Lin, J., Khade, R., and Li, Y. (2012). Rotation-invariant

similarity in time series using bag-of-patterns repre-

sentation. Journal of Intelligent Information Systems,

39(2):287–315.

Lorenz, D. A comparison of denoising methods for one

dimensional time series.

Marteau, P.-F. (2009). Time warp edit distance with stiff-

ness adjustment for time series matching. IEEE Trans-

actions on Pattern Analysis and Machine Intelligence,

31(2):306–318.

Matuszyk, P., Castillo, R. T., Kottke, D., and Spiliopoulou,

M. A comparative study on hyperparameter optimiza-

tion for recommender systems.

Radu, V., Katsikouli, P., Sarkar, R., and Marina, M. K.

(2014). A semi-supervised learning approach for ro-

bust indoor-outdoor detection with smartphones. In

Proceedings of the 12th ACM Conference on Embed-

ded Network Sensor Systems, pages 280–294. ACM.

Rakthanmanon, T. and Keogh, E. (2013). Fast shapelets:

A scalable algorithm for discovering time series

shapelets. In proceedings of the 2013 SIAM Interna-

tional Conference on Data Mining, pages 668–676.

SIAM.

Ratanamahatana, C. A. and Keogh, E. (2004). Everything

you know about dynamic time warping is wrong. Cite-

seer.

Sch

afer, P. (2015). The boss is concerned with time series

classiﬁcation in the presence of noise. Data Mining

and Knowledge Discovery, 29(6):1505–1530.

Tapinos, A. (2013). Time Series Data Mining In Systems

Biology. PhD thesis, The University of Manchester

(United Kingdom).

Wan, J., O’grady, M. J., and O’hare, G. M. (2015). Dy-

namic sensor event segmentation for real-time activ-

ity recognition in a smart home context. Personal and

Ubiquitous Computing, 19(2):287–301.

Xi, X., Keogh, E., Shelton, C., Wei, L., and Ratanama-

hatana, C. A. (2006). Fast time series classiﬁcation us-

ing numerosity reduction. In Proceedings of the 23rd

international conference on Machine learning, pages

1033–1040. ACM.

Yazdansepas, D., Saroha, N., Ramaswamy, L., and

Rasheed, K. Towards efﬁcient & real-time human ac-

tivity recognition using wearable sensors: A shapelet-

based pattern matching approach.

HEALTHINF 2020 - 13th International Conference on Health Informatics