Evaluating Sensor Placement and Feature Importance for Hurling

Movement Classification

Chloe Leddy

, Richard Bolger

, Paul J. Byrne

, Sharon Kinsella

and Lilibeth Zambrano

Department of Aerospace and Mechanical Engineering, South East Technological University,

Kilkenny Road Campus, Carlow, Ireland

Department of Sport and Exercise Science, South East Technological University, Cork Road Campus, Waterford, Ireland

Department of Health and Sport Sciences, South East Technological University, Kilkenny Road Campus, Carlow, Ireland

Keywords: Machine Learning, Signal Processing, Inertial Measuring Unit, Feature Importance, Team Sport.

Abstract: Human Activity Recognition (HAR) involves recognising and classifying human activities from data

collected by sensors through machine learning (ML) techniques. The assessment of athletic movement via

HAR has benefited sport performance analysis by identifying technical and tactical performance indicators.

Hurling is a dynamic stick and ball invasion team sport that involves high-impact movements. Sensor

placement and feature selection in HAR tasks impact the classification accuracy of the ML model during

testing and training. This study aims to determine the optimal inertial measuring unit (IMU) sensor placement

for recognizing hurling movements and to identify the most important features for accurate classification.

Time-domain and frequency-domain features of accelerometer data were computed and were used to train

and test three classification models: Support Vector Machine (SVM), Random Forest (RF) and k-Nearest

Neighbour (k-NN). A RF model achieved the highest mean accuracy in the recognition of four hurling specific

movements, for sensors located at the forearm (86%) and the thigh (84%). Features extracted from the z-axis,

specifically zero crossing rate (ZCR), standard deviation (STD), and root mean square (RMS) were most

discriminative in classifying hurling sport movements with a RF model using a forearm-mounted IMU.

1 INTRODUCTION

Hurling is a stick and ball team sport that is

predominantly played in Ireland and involves high

intensity, intermittent activity (Mullane et al., 2018).

The sport involves a multitude of advanced technical

skills and requires the proficient use of a stick (Hurl)

to control and strike a ball (sliotar) at high velocities

(Leddy et al., 2023). Hurling encompasses a broad

range of skills and physiological considerations such

as explosive power, striking a ball in the air, jumping,

and sprinting (Collins et al., 2022). Successful

performance outcomes in hurling match-play are

linked with an understanding and knowledge of

physical and physiological demands (Keane et al.,

2021). Activity monitoring of team sports leads to

https://orcid.org/0000-0002-8196-7648

https://orcid.org/0000-0003-2836-0166

https://orcid.org/0000-0002-4976-9131

https://orcid.org/0000-0001-9051-4467

https://orcid.org/0000-0002-0392-7309

increased knowledge of physical and physiological

in-game demands and assists in performance

profiling, training prescription and reduces the

likelihood of injury (Ribeiro et al., 2020). The

increased desire to understand sports motion has led

to motivated research in sports activity recognition

which has examined the frequency, intensity,

duration, and type of activity performed during

competitive and training events (Pfeiffer et al., 2023;

Ren et al., 2016).

Monitoring and automatic recognition of physical

activities is often referred to as a human activity

recognition (HAR) task (Bulling et al., 2014). HAR is

a challenging time-series task that has been used in

team sports such as Australian football (Cust et al.,

2021), field hockey (Shahar et al., 2020) and cricket

Leddy, C., Bolger, R., Byrne, P., Kinsella, S. and Zambrano, L.

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation.

DOI: 10.5220/0012904000003828

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 12th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2024), pages 34-45

ISBN: 978-989-758-719-1; ISSN: 2184-3201

(McGrath et al., 2021) to detect and identify actions

of athletes. HAR in sports applications has been

shown to be beneficial for measuring training volume

(Hendry et al., 2020), player performance evaluation

and assessing biomechanical factors of sports

movement (McDevitt et al., 2022; Roslan & Ahmad,

2020). The continuous and coupled developments in

technology and artificial intelligence (AI) over the

past two decades has enabled sports activity

recognition that is robust, accurate and increasingly

automated (Chmait, & Westerbeek, 2021). An inertial

measuring unit (IMU) is a sensor system that

combines a gyroscope, accelerometer and often a

magnetometer for measurements of angular velocity,

acceleration, and orientation, and are commonly used

as data inputs for classification of human movement

(Kranzinger et al., 2023). A review of wearable

technology in sport reported that IMU and its

subcomponent accelerometer were the main

keywords featured in 2568 research articles which

were indexed in the Social Sciences Citation Index

(SSCI) or the Science Citation Index Expanded (SCI-

E) (Seçkin et al., 2023). Tri-axial accelerometers are

particularly suited to activity recognition due to their

ability to measure acceleration proportional to

external force allowing for measurements of dynamic

movements reflecting changes in activity intensity

and frequency (Twomey et al., 2018).

The HAR framework is depicted in Figure 1.

Typically, the following steps are involved: data

acquisition, data preprocessing (signal processing and

segmentation techniques), feature extraction,

classification through AI techniques and performance

evaluation (Bento et al., 2022). The application of AI

in the form of a machine learning (ML) or deep

learning (DL) model for classification of activities

and an associated performance evaluation of said

model is an integral component of the HAR

framework (Kulsoom et al., 2022). Studies have

shown that each step of the ML modelling process are

iterative, and classification accuracy depends on the

specific characteristics of the movement being

analysed (Gil-Martín et al., 2020).

Traditional supervised ML models such as Support

Vector Machine (SVM), Random Forest (RF) and k-

nearest neighbour (k-NN) have been extensively

employed for activity recognition tasks based on

accelerometer data (Slim et al., 2019) and in the

classification of human motion data based on IMUs

in sports (Kranzinger et al., 2023). SVM is widely

reported in the literature for the classification of

spatiotemporal features into activity categories in

sensor-based sport activity recognition (Cust et al.,

2021). Naïve bayes (NB), Decision Tree (DT) and

Convolutional Neural Networks are also commonly

applied ML models in the HAR research area (Pajak

et al., 2022). The k-NN algorithm has demonstrated

strong performance in the classification of human

activities (Mohsen et al., 2022). A weighted k-NN

model achieved 82.5 ± 4.75% accuracy in the

prediction and classification of performance

indicators attributed to the shooting score in archery

(Muazu Musa et al., 2019). In HAR research, it is

common to apply several ML models to determine the

best fit for the recognition and classification of the

activities (Preatoni et al., 2020), as the classification

performance of the ML model will be dependent upon

the characteristics of the dataset under investigation.

Figure 1: Sports Activity Recognition Framework.

The data pre-processing is the second step in sport

activity recognition where the data is filtered and

segmented to define activity boundaries through

techniques such as overlapping windows. The

statistical and mathematical features are then

extracted from each window to prepare for use with

ML models. Feature extraction captures the relevant

information required to differentiate activities

represented by the sensor signals during sports

movement. Feature premutation refers to the

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation

assessment of the impact of feature relevance on a

model’s performance on given datasets (Vallance et

al., 2020). For HAR tasks from IMU data, features

are typically extracted from the time-domain and

frequency-domain (Gomaa & Kamas, 2023; Rosati et

al., 2018). Features extracted from the time domain

reveal statistical information about the signal and are

the most utilised method in HAR tasks (Rosati et al.,

2018). Studies have shown that time-domain features

may be sufficient to classify an activity class (Chong

et al., 2021). Frequency-domain features reveal

information about the signal’s periodicity, such as

underlying oscillations which is beneficial in the

recognition of activities with distinct periodic

patterns (Dehkordi et al., 2020). To obtain high

performance accuracy, the input features must be

representative of the movement being analyzed. The

selection of appropriate features has been of interest

in HAR research (Allik et al., 2019; Bennasar et al.,

2022). Studies have found that the number of features

(Brzostowski and Szwach, 2018) in addition to the

type of features (McGrath et al., 2021) have an

influence on the classification performance of ML

models in sport activity recognition tasks.

Performance evaluation metrics are quantitative

methods used at the end of the HAR pipeline to

determine the effectiveness of the classification

model. Performance is generally assessed using

accuracy, which is a measure of the number of correct

predictions divided by the total number of predictions

and is derived from a confusion matrix (Ward et al.,

2011). Accuracy may be over predicted when the

classes in a dataset are imbalanced, or if there is

insufficient information on the instances of false

positives and false negatives. Precision and recall are

metrics which can be used to supplement accuracy

when evaluating a model’s classification

performance. Precision focuses on minimizing false

positives, while recall aims on minimizing false

negatives (Ward et al., 2011). The classification

accuracy is highly dependent on the incoming data,

and as such the sensor location should correspond

with the movement being analyzed. For example,

ankle-mounted IMUs produced high accuracy of 80-

83%, in the in-situ classification of Australian

football kick types (Cust et al., 2021). An

investigation into the optimal sensor placement for

badminton found that a sensor placed on the bottom

of the rackets grip provided the best recognition

accuracy when examining stroke types (Steels et al.,

2020). Other studies investigating the influence of

sensor location on the recognition of complex

movements found that a combination of sensors

achieved the highest performance of 96.7% accuracy

(Shahar et al., 2020) and 97.6% (Preatoni et al., 2020)

respectively. The placement of sensors at varying

body segments and sensor combinations should be

explored to determine classification accuracy for

optimised sport activity recognition (Xia & Sugiura,

2021).

Extensive research has been conducted on the use

of accelerometer data and ML techniques for accurate

sport activity recognition and classification (Cust et

al., 2019; Pfeiffer et al., 2023). However, to the best

of current knowledge, the investigation of HAR to

classify hurling movements has not yet been

conducted. This study aims to determine the optimal

IMU sensor placement for recognizing hurling

movements and to identify the most important

features for accurate classification using three ML

models; SVM, RF, k-NN.

2 METHODS

2.1 Data Collection

A total of four hurling specific sport activities were

performed for 1-min each, over an 8-min period, with

1-min intervals of rest between each activity type.

Five hurling players (age 22.0 ± 5.61 years; height

178.4 ± 5.64 cm; body mass 83.6 ± 17.73 kg) with an

average training age (Myer at al., 2013) of 16.8±4.09

years participated in this study. The activities

included (1) jab lift, (2) overhead catch, (3) soloing,

and (4) striking. A description of these activities is

detailed in Table 1 below. Ethical approval was

obtained from the South East Technological

University Research Ethics Committee

(Ethical approval code: 332).

The dataset analysed in this study was a primary

dataset collected utilising the Xsens Motion

Visualization and Navigation (MVN) link inertial

measuring system (Movella Technologies B.V.,

Enschede, Netherlands). The Xsens MVN link is a 3D

motion capture system which consists of 17 IMUs

that are wired and tightly affixed to body segments in

a Lycra suit. The data used in this study was

accelerometer derived data obtained from the IMUs

located at the forearm, and right upper leg, which will

be referred to hereafter as ‘thigh’. The forearm sensor

is positioned on the dorsal (posterior) side of the

forearm, just above the wrist joint. The thigh sensor

is positioned a few centimetres above the mid-thigh,

or on the iliotibial band on the external side of the leg

(Movella, 2022). The 23 segments of the kinematic

model are defined according to the international

Society of Biomechanics (ISB) recommendations,

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

and detailed sensor placement information can be

found in the MVN user manual (Movella, 2024).

The accelerometer range for Xsens MVN link

system is MTx: ± 160 m/s2 (16 g) MTw: ±160m/s2

(16g) (Movella, 2024). The movements were

simultaneously recorded by a Panasonic HX-WA20

camera with a resolution rate of 1920px X 1080px and

a frame rate of 30fps. The cameras, which were

mounted on stationary tripods, were used to establish

ground truth. All statistical and visual analysis were

conducted in a Python environment (Python, 3.8.12)

Table 1: Hurling Activities Description.

Activity Description

Jab Lift

Player slides the hurl under the

ball and scoops it into the hand in

one swift action.

Overhead

catch

Player positions themselves under

the flight path of the sliotar,

anticipating its descent. The

player then jumps into the air off

one leg, bending the opposite leg,

and the ball is caught with a

cupped hand.

Soloing

Player balances the sliotar on top

of the hurl as they take steps. The

hurl is held out in front of the

player.

Striking

Player positions their dominant

hand at the top of the handle and

their non-dominant hand further

down the handle. The hurl is

swung above the head until it

meets contact with the sliotar,

where it is struck.

2.2 Data Preprocessing

IMU sensors measuring dynamic movements may be

susceptible to noise and drift in the signals due to

magnetic disturbances and offset errors owing to the

participants’ unintentional shaking or movement.

These movements may present as slight and

potentially repetitive fluctuations which distort the

signal, affecting the quality of the movement data

captured and the classification performance of the

associated machine learning model (de Cheveigné &

Nelken, 2019).

Correspondingly, the application of

lowpass filtering is an integral component of the

activity recognition framework following on from

data acquisition (Hsu et al., 2018). The purpose of a

filter is to remove interference noise, miscellaneous

signal fluctuations, and low-frequency components

(Yin et al., 2021). A low pass filter only allows lower

signal frequencies that are below its cutoff frequency

to pass through, while attenuating all signals above

the cut off frequency, effectively extracting the useful

components of the signal relating to physical activity

which lie within a specific frequency range (Shouran

& Elgamli, 2020). Fridolfsson et al. (2019) suggested

that accelerations related to human movements are

typically found between 1 and 10 Hz. The comparison

between filtered and unfiltered accelerometer data (x,

y, z axes) recorded from the forearm mounted IMU

of a participant during 60 seconds of striking is

displayed in Figure 2.

Figure 2: Comparison of Unfiltered (blue line) and Filtered

(red line) Acceleration Data for Striking Movement

obtained by the Forearm IMU sensor.

In this study, a low-pass 4th order Butterworth

filter with a cut-off frequency of 10 Hz using a

second-order filter two times to the time series was

implemented to smooth the signal by attenuating

frequencies higher than 10 Hz. A fourth-order

Butterworth filter is commonly used in

biomechanical applications (Crenna et al., 2021) and

in motion recognition (Liu et al., 2022). Additionally,

a moving average filter with a window size of 5

samples was applied to smooth the data. Once the

data has been filtered, it is segmented into

overlapping windows to facilitate feature extraction

and model training (Cero Dinarević et al., 2019). The

optimal window size and overlap depends on the

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation

specific characteristics and application of the dataset.

Bonomi et al. (2009) found that reducing the segment

size decreased the machine learning classification

accuracy for physical activity recognition.

Conversely, Brzostowski and Szwach (2018)

reported that increasing the window size to 80

samples improved the classification performance of

k-NN and Logistic Regression for classification of

stroke type in tennis.

Figure 3: Sport Activity Recognition Framework applied in

this research.

For this study, a sliding window approach of 2.5s

with a 50% overlap was applied. The 2.5s window

size demonstrates the most accurate detection

performance for common sports activity for

supervised machine learning models (Ghazali et al.,

2018). The features were extracted from each window

segment in both the time domain and frequency

domain for the x, y, and z axes. The extracted features

are detailed as follows: mean, standard deviation

(SD), root mean square (RMS), skewness, kurtosis,

zero crossing rate (ZCR), dominant frequency, and

total power. These features have been widely adopted

in previous studies examining accelerometer-based

activity recognition (Gomaa & Khamis, 2023). The

time-domain features in this study are simple features

extracted through basic statistical analysis providing

characteristics of the signal over time and are highly

effective in discriminating activities from

accelerometer signal (Erdaş et al., 2016). Simple

statistical features have shown outstanding

classification accuracy in differentiating static and

dynamic activities (Coelho et al., 2022). Frequency-

domain features are extracted through spectral

analysis and compliment time-domain features

(Erdaş et al., 2016) for robust feature representation

(Nguyen et al., 2021). The hurling activities

examined in this research exhibit a combination of

rotational and linear movements with varying

intensities and as such, features with high

discriminative abilities are required to accurately

recognise and classify different activities. For

example, striking movements often involve rapid,

intense movements leading to greater variability and

distinct skewness and kurtosis values compared to

more uniform activities. Additionally, frequency

domain features reveal underlying oscillations and

energy distributions in the data. Activities such as

jogging (soloing) and jumping (Overhead catch) have

indicative frequency components that can be detected

through these features.

The train/test split of the data was 80/20%, where

80% of the data was used to train the machine learning

models (SVM, RF, and k-NN), and the remaining 20%

was used for performance evaluation. In this study, a

5-fold cross validation was implemented. K-Fold cross

validation involves randomly partitioning the data in k-

equal subsets, training the data on k-1 subset, and using

a different fold for testing (Dehghani et al., 2019). This

process is repeated k times and when the k iterations

are completed, performance metrics such as accuracy,

precision and recall are calculated by averaging the

results of all k folds. The sport activity recognition

framework implemented in this study is shown in

Figure 3.

3 RESULTS AND DISCUSSION

Each of the five participants performed each activity

for 1 minute, which corresponds to 240 s at a

sampling frequency of 240 Hz/s. For each activity,

14400 samples were collected per axis, resulting in

172800 samples per participant across all three axes.

The data was segmented into fixed sized windows of

2.5 s with a 50% overlap this translates to 600

samples per window with an overlap of 300 samples,

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

this segmentation yielded 940 windows per sensor.

Each window was processed to extract various time-

domain and frequency-domain features. Three

supervised classification models, specifically SVM,

RF and k-NN were compared to determine the best

performance in terms of mean accuracy (A), mean

precision (P), and mean recall (R), these results are

displayed in Table 2. Figure 4 presents the confusion

matrices, while Tables 3 and 4 provide a detailed

breakdown of the TPs, TNs, FPs, and FNs for our

classification models using the forearm and thigh

sensors. Permutation feature importance with a cut-

off threshold of 0.05 was conducted to determine the

importance of different features across the time and

frequency domain for each of the sensor locations in

predicting the hurling activity classes.

Figure 4: The confusion matrix for Random Forest model

using the forearm sensor for the detection of four Hurling

sport specific activities.

Table 2: Performance metrics of SVM, RF, and k-NN

Models for the recognition and classification of four hurling

specific activities (A = mean accuracy, P = mean precision,

and R = mean recall. These metrics summarise the model

performance).

Sensor

Location

Model Mean A Mean P Mean R

Forearm SVM

0.848 0.850 0.848

0.863

0.866

0.864

k-NN 0.741

0.743

0.740

Thigh SVM 0.815

0.816 0.814

RF 0.842

0.845 0.844

k-NN 0.659

0.657 0.657

The confusion matrices from the RF model for the

forearm sensor and thigh sensor are displayed in

Figure 4 and Figure 5, respectively. The high values

on the diagonal (correctly classified instances)

suggest that the RF model is performing favourably

for the classification of hurling activities from both

sensor locations.

The analysis of two sensor locations revealed the

RF model as the best-performing classifier with mean

accuracy of 86% for the forearm, and 84% for the

thigh respectively.

Figure 5: The confusion matrix for Random Forest model

using the thigh sensor for the detection of four Hurling sport

specific activities.

Table 3: Mean True Positives (TPs), True Negatives (TNs),

False Positives (FPs) and False Negatives (FNs) for SVM,

RF and k-NN classifiers on Forearm Sensor Data.

Model Class Mean

TPs

Mean

TNs

Mean

FPs

Mean

FNs

SVM 0 42.4 ±

6.46

136.0 ±

5.96

5.0 ±

2.0

4.6 ±

1.35

1 43.2 ±

3.65

137.4 ±

4.07

3.6 ±

3.2

3.8 ±

1.93

2 39.4 ±

7.00

133.2 ±

7.93

7.8 ±

1.16

7.6 ±

2.87

3 34.6 ±

3.84

129.0 ±

6.29

12.0 ±

3.82

12.4 ±

2.80

RF 0 43.4 ±

5.12

136.4 ±

6.97

4.6 ±

1.01

3.6 ±

1.85

1 41.4 ±

4.49

138.2 ±

2.71

2.8 ±

2.22

5.6 ±

2.41

2 40.2 ±

6.67

135.0 ±

7.18

6.0 ±

1.89

6.8 ±

2.56

3 37.4 ±

2.05

128.8 ±

6.01

12.2 ±

2.99

6.6 ±

2.57

k-NN 0 40.0 ±

4.09

133.6 ±

4.84

7.4 ±

3.38

7.0 ±

2.09

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation

able 4: Mean True Positives (TPs), True Negatives (TNs),

alse Positives (FPs) and False Negatives (FNs) for SVM,

F and

-NN classifiers on Forearm Sensor Data.(cont.)

1 36.6 ±

5.78

133.2 ±

4.53

7.8 ±

1.46

10.4 ±

3.32

2 37.4 ±

7.14

121.2 ±

9.62

19.8 ±

5.15

9.6 ±

2.65

3 25.4 ±

3.97

127.4 ±

6.77

13.6 ±

3.44

21.6 ±

3.26

Table 3 and 4 summarize the TP (True Positives),

TN (True Negatives), FP (False Positives), and FN

(False Negatives) values for each classifier across both

forearm (Table 3) and thigh sensors (Table 4). TP are

the number of positive cases correctly identified as

positive; the activity is classified accurately. TN are the

number of negative cases correctly identified as

negative; the activity is classified accurately. FP are the

number of negative cases incorrectly identified as

positive; the activity is misclassified. FN are the

number of positive cases incorrectly identified as

negative; the activity is misclassified. (Bennasar et al.,

2022). For both sensor locations, SVM and RF models

showed a decline in performance as denoted by a

decrease in TPs and TNs, and an increase in FPs and

Table 5: Mean True Positives (TPs), True Negatives (TNs),

False Positives (FPs) and False Negatives (FNs) for SVM,

RF and k-NN classifiers on Thigh Sensor Data.

Model Class Mean

TPs

Mean

TNs

Mean

FPs

Mean

FNs

SVM 0 38.6 ±

2.24

134.2 ±

3.81

6.8 ±

2.63

8.4 ±

1.35

1 39.2 ±

5.6

130.8 ±

3.54

10.2 ±

1.93

7.8 ±

2.03

2 39.0 ±

7.42

135.4 ±

7.86

5.6 ±

2.72

8.0 ±

2.09

3 36.6 ±

6.49

129.0 ±

7.12

12.0 ±

2.75

10.4 ±

2.57

RF 0 40.2 ±

2.92

135.4 ±

4.33

6.0 ±

2.09

6.8 ±

1.72

1 42.6 ±

3.92

130.0 ±

2.52

11.0 ±

4.60

4.4 ±

2.93

2 39.8 ±

5.91

137.6 ±

6.46

3.4 ±

3.07

7.2 ±

2.71

3 35.8 ±

4.99

131.8 ±

6.24

9.2 ±

3.12

11.2 ±

2.31

k-NN 0 35.6 ±

3.07

125.0 ±

5.32

16.0 ±

2.52

11.4 ±

3.07

1 31.8 ±

3.12

119.6 ±

3.72

21.4 ±

4.63

15.2 ±

2.78

2 32.6 ±

6.56

130.0 ±

7.89

11.0 ±

3.40

14.4 ±

4.96

3 24.0 ±

7.66

125.4 ±

6.18

15.6 ±

2.05

23.0 ±

2.96

FNs, particularly in class 3. However, both SVM and

RF present as generally reliable with high mean TPs

and TNs values throughout. RF was revealed as the

better-performing classifier, showing notable

performance with the forearm sensor. In contrast, the

k-NN model had a significant drop in performance

across both sensor locations with higher FPs and FNs

lower TPs and TNs.

The importance of features based on their impact

on predictive performance of a RF model was

calculated for each of the sensor locations through

permutation feature importance, as displayed in

Figure 5 and Figure 6. The features that demonstrated

the greatest importance for the forearm mounted IMU

with the RF classifier were ZCR, STD, and RMS in

the z-axis, whereas total power from the x-axis, and

mean from the y axis were of little predictive power.

The most important features for the thigh mounted

IMU were STD from the y axis, mean from the z-axis,

and RMS from the y-axis. Features extracted from the

x-axis, specifically RMS and total power were of least

predictive power. Time-domain features, such as std,

mean and rms are particularly effective for capturing

the magnitude and variability of lower body

accelerations in activities such as jumping (overhead

catch) and jogging (soloing). Frequency domain

features, such as dominant frequency, are particularly

relevant for activities characterised by complex and

rhythmic motions, which are common in striking

activities. These features help in identifying activities

with prominent rhythmic components.

This study shows that the RF model achieved the

highest performance as denoted by the mean

accuracy, precision and recall in both sensor

locations. This result is synonymous with the work of

Hölzemann & Van Laerhoven (2018) who examined

the performance of ML models for classification of

basketball activities, reporting that a RF model

achieved the greatest mean accuracy of 87.5%

outperforming a k-NN model. Similarly, a RF

outperformed a k-NN and SVM models for the

classification of human daily activities, and in these

experiments, the highest accuracies of a RF model

were achieved when the classifier was fed with time-

domain features only (Erdaş et al., 2016) and a

combination of time-domain and frequency-domain

features (Nurwulan & Selamaj, 2020).

Random Forests are ensemble methods that

combine multiple random decision trees (in this study

the model consisted of 100 decision trees), each tree

is trained on a random subset of the data (Breiman,

2001). Thus, the random sampling and aggregation of

predictions results in a classifier that is scalable,

efficient and robust to overfitting, enabling it to

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

capture a broad range of patterns and characteristics

from complex sporting activities. RF identifies the

most relevant features, and research has shown that

features extracted from the z-axis are of most

importance (Casale et al., 2019).

Figure 6: Permutation Feature Importance of Thigh

Mounted IMU Features for Random Forest Classifier in

Classifying four Hurling Movements.

Figure 7: Permutation Feature Importance of Forearm

Mounted IMU Features for Random Forest Classifier in

Classifying four Hurling Movements.

A novel activity recognition system which

combined nonparametric weighted feature extraction

(NWFE), principal component analysis (PCA) and

least squared support vector machine (LS-SVM) was

introduced by Hsu et al. (2018) for the recognition of

10 daily human activities and 11 sport activities and

reported an overall correct classification rate (CCR)

of between 98.33 – 99.55%. Research in HAR is

underpinned by the “No Free Lunch” theorem which

explains that there is no universal best fit algorithm

(Wolpert, & Macready, 1997). This is particularly

evident in HAR research whereby differences are

evident in data collection specifications, data pre-

processing techniques, machine learning modelling

and evaluation. As aforementioned, every step of the

HAR pipeline is iterative and should be considered in

alignment with the characteristics of the movement

being analysed.

The optimal sensor placement has been of

significant interest in HAR research (Davoudi et al.,

2021; Steels et al., 2020). In this study, it was

demonstrated that a single IMU mounted at the right

forearm coupled with a RF model achieved a

marginally higher accuracy compared to a thigh

mounted IMU sensor for the recognition and

classification of hurling activities. Shahar et al.

(2020) examined the influence of sensor combination

and location for activity recognition in field hockey

and reported that a left wrist mounted sensor achieved

86.2% accuracy with a cubic SVM compared to other

single sensor locations (waist, right wrist, and chest).

However, the highest accuracy (96.7%) in these

experiments was achieved when all 4 sensor locations

were combined.

By combining several sensors from varying body

segments, the classification performance of an ML

model may be improved (Davoudi et al., 2021), but

there are drawbacks including labour intensive post-

processing and increased computational load.

Moreover, considering the importance of ecological

validity in sport science, HAR research should take

place, where possible, in the athletes’ natural sporting

environment. Using multiple sensors in these training

and competitive environments may prove to be

cumbersome for athletes, and in such scenarios a

single body-mounted sensor may be more practical.

Permutation feature importance is a technique

used to determine the importance of different features

in a predictive model. It works by randomly shuffling

the value of each feature and measuring the resultant

decrease in the model’s performance, such as

prediction accuracy and area under the curve (AUC)

(Vallance et al., 2020). The larger the drop in ML

performance, the more important that feature is. The

analysis of feature importance in this study

highlighted that simple time-domain features,

particularly those extracted from the z-axis, were

generally most relevant. Other research on significant

features for HAR using tri-axial accelerometers

reported that simple time domain features were of

most significance (Bennasar et al., 2022).

comparison of two feature sets for HAR, showed that

a feature set comprised exclusively of time-domain

features achieved a performance of 96.7% compared

to a more complex feature set, containing both time-

domain and frequency-domain derived features

which obtained a slightly higher performance of

97.1% (Rosati et al., 2018). The previous research

suggest that time-domain features can be highly

effective for HAR tasks, but the addition of frequency

domain features can reveal underlying patterns and

oscillations in the data and contribute to the accurate

classification of complex sports movements

(Dehkordi et al., 2020; Tran et al., 2014).

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation

4 CONCLUSION

In this study, a sports activity recognition framework

is proposed for the classification of four hurling sport

specific movements. Accelerometer data were

collected from IMUs mounted on the forearm and

thigh of five hurling athletes. The performance of

SVM, RF and k-NN models for the recognition and

classification of Hurling activities was assessed for

each sensor location. Additionally, the most relevant

features for activity classification were examined

though permutation feature importance. According to

the study results, the RF achieved the best result in

both cases represented by a mean accuracy of 86% for

the forearm sensor, and 84% for the thigh sensor,

respectively. The analysis revealed that time-domain

features extracted from the y-axis and z-axis were of

most importance for the thigh sensor in their

contribution to the RF model’s predictive power.

Similarly, for the forearm sensor, time-domain

features extracted from the z-axis were most

important, specifically ZCR.

This study demonstrates that dynamic field sports

involving non-cyclical movements, such as hurling,

are amenable to human activity recognition research.

Traditional machine learning models, namely SVM,

RF and k-NN showed favourable results, as

demonstrated by mean accuracies between 74% -

86% for the forearm sensor location, and 65% - 84%

for the thigh sensor location. Future research in this

field may consider combining features from different

sensor locations for increased event detection and

model generalisability in future scenarios. But if one

sensor is preferred, a sensor mounted at the forearm

for recognition and classification of hurling activities

is recommended.

The limited sample size of 5 participants reflects

the specific inclusion criteria for hurling players and

the availability of participants at the time of testing.

The small sample size may affect the generalisability

and interpretation of the findings. However, given

that this is the first research examining activity

recognition in the sport of hurling, this research is

exploratory in nature and provides valuable insights.

Moreover, the methodology outlined in this research

did not include hyperparameter tuning for the selected

models. As a result, this may have affected the

generalisability and accuracy of the models. Future

research may consider hyperparameter tuning

techniques such as Grid Search or Bayesian

Optimization to enhance the performance and

generalisability of SVM, k-NN, and RF models.

REFERENCES

Allik, A., Pilt, K., Karai, D., Fridolin, I., Leier, M., &

Jervan, G. (2019). Optimization of physical activity

recognition for real-time wearable systems: Effect of

window length, sampling frequency and number of

features. Applied Sciences, 9(22), Article 4833.

https://doi.org/10.3390/app9224833

Bennasar, M., Price, B. A., Gooch, D., Bandara, A. K., &

Nuseibeh, B. (2022). Significant Features for Human

Activity Recognition Using Tri-Axial

Accelerometers. Sensors (Basel, Switzerland), 22(19),

7482. https://doi.org/10.3390/s22197482

Bento, N., Rebelo, J., Barandas, M., Carreiro, A. V.,

Campagner, A., Cabitza, F., & Gamboa, H. (2022).

Comparing Handcrafted Features and Deep Neural

Representations for Domain Generalization in Human

Activity Recognition. Sensors (Basel,

Switzerland), 22(19), 7324. https://doi.org/10.3390/

s22197324

Bonomi, A. G., Goris, A. H., Yin, B., & Westerterp, K. R.

(2009). Detection of type, duration, and intensity of

physical activity using an accelerometer. Medicine and

science in sports and exercise, 41(9), 1770–1777.

https://doi.org/10.1249/MSS.0b013e3181a24536

Breiman, L. (2001). Random Forests. Machine Learning,

45, 5–32. https://doi.org/10.1023/A:1010933404324

Brzostowski, K., & Szwach, P. (2018). Data fusion in

ubiquitous sports training: Methodology and

application, Wireless Communications and Mobile

Computing, 1-14. https://doi.org/10.1155/2018/8180296

Bulling, A., Blanke, U., & Schiele, B. (2014). A tutorial on

human activity recognition using body-worn inertial

sensors. ACM Computing Surveys, 46, 33:1-33:33.

https://doi.org/10.1145/2499621

Casale, P., Pujol, O., & Radeva, P. (2011). Human Activity

Recognition from Accelerometer Data Using a

Wearable Device. In: Vitrià, J., Sanches, J.M.,

Hernández, M. (eds) Pattern Recognition and Image

Analysis. IbPRIA 2011. Lecture Notes in Computer

Science, vol 6669. Springer, Berlin, Heidelberg.

https://doi.org/10.1007/978-3-642-21257-4_36

Cero Dinarević, E., Baraković Husić, J., & Baraković, S.

(2019). Step by Step Towards Effective Human

Activity Recognition: A Balance between Energy

Consumption and Latency in Health and Wellbeing

Applications. Sensors (Basel, Switzerland), 19(23),

5206. https://doi.org/10.3390/s19235206

Chmait, N., & Westerbeek, H. (2021). Artificial

Intelligence and Machine Learning in Sport Research:

An Introduction for Non-data Scientists. Frontiers in

sports and active living, 3, 682287. https://doi.org/10.

3389/fspor.2021.682287

Chong, J., Tjurin, P., Niemelä, M., Jämsä, T., & Farrahi, V.

(2021). Machine-learning models for activity class

prediction: A comparative study of feature selection

and classification algorithms. Gait & Posture, 89, 45-

53. https://doi.org/10.1016/j.gaitpost.2021.06.017

Coelho, Y. L., Nguyen, B., Santos, F. A., Krishnan, S., &

Bastos-Filho, T. F. (2022). A lightweight model for

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

human activity recognition based on two-level

classifier and compact CNN model. In T. F. Bastos-

Filho, E. M. de Oliveira Caldeira, & A. Frizera-Neto

(Eds.), XXVII Brazilian Congress on Biomedical

Engineering: CBEB 2020 (Vol. 83, pp. 1944-1953).

Springer, Cham. https://doi.org/10.1007/978-3-030-

70601-2_276

Collins, K., Reilly, T., Malone, S., Keane, J., & Doran, D.

(2022). Science and Hurling: A review. Sports, 10(8),

118. https://doi.org/10.3390/sports10080118

Cust, E. E., Sweeting, A. J., Ball, K., & Robertson, S.

(2019). Machine and deep learning for sport-specific

movement recognition: a systematic review of model

development and performance. Journal of sports

sciences, 37(5), 568–600. https://doi.org/10.1080/0264

0414.2018.1521769

Cust, E. E., Sweeting, A. J., Ball, K., & Robertson, S.

(2021). Classification of Australian football kick types

in-situation via ankle-mounted inertial measurement

units. Journal of sports sciences, 39(12), 1330–1338.

https://doi.org/10.1080/02640414.2020.1868678

Crenna, F., Rossi, G. B., & Berardengo, M. (2021).

Filtering biomechanical signals in movement analysis.

Sensors, 21(13), 4580. https://doi.org/10.3390/s211345

de Cheveigné, A., & Nelken, I. (2019). Filters: When, Why,

and How (Not) to Use Them. Neuron, 102(2), 280–293.

https://doi.org/10.1016/j.neuron.2019.02.039

Dehkordi, M. B., Zaraki, A., & Setchi, R. (2020). Feature

extraction and feature selection in smartphone-based

activity recognition. Procedia Computer Science, 176.

2655-2664. https://doi.org/10.1016/j.procs.2020.09.301

Davoudi, A., Mardini, M. T., Nelson, D., Albinali, F.,

Ranka, S., Rashidi, P., & Manini, T. M. (2021). The

Effect of Sensor Placement and Number on Physical

Activity Recognition and Energy Expenditure

Estimation in Older Adults: Validation Study. JMIR

mHealth and uHealth, 9(5), e23681. https://doi.org/10.

2196/23681

Dehghani, A., Glatard, T., & Shihab, E. (2019). Subject

Cross Validation in Human Activity

Recognition. ArXiv, abs/1904.02666.

https://doi.org/10.48550/arXiv.1904.02666

Erdaş, Ç. B., Atasoy, I., Açıcı, K., & Oğul, H. (2016).

Integrating features for accelerometer-based activity

recognition. Procedia Computer Science, 98, 522-527.

https://doi.org/10.1016/j.procs.2016.09.070

Fridolfsson, J., Börjesson, M., Buck, C., Ekblom, Ö.,

Ekblom-Bak, E., Hunsberger, M., Lissner, L., &

Arvidsson, D. (2019). Effects of Frequency Filtering on

Intensity and Noise in Accelerometer-Based Physical

Activity Measurements. Sensors (Basel,

Switzerland), 19(9), 2186. https://doi.org/10.3390/s

19092186

Ghazali, N. F., As’ari, M. A., Shahar, N., & Latip, H. F. M.

(2018). Investigation on the effect of different window

size in segmentation for common sport activity. In 2018

International Conference on Smart Computing and

Electronic Enterprise (ICSCEE) (pp. 1-7). Shah Alam,

Malaysia. IEEE. https://doi.org/10.1109/ICSCEE.

2018.8538429

Gil-Martín, M., San-Segundo, R., Fernández-Martínez, F.,

& de Córdoba, R. (2020). Human activity recognition

adapted to the type of movement. Computers &

Electrical Engineering, 88, 106822. https://doi.org/10.

1016/j.compeleceng.2020.106822

Gomaa, W., & Khamis, M.A. (2023) A perspective on

human activity recognition from inertial motion

data. Neural Computing & Application, 35, 20463–

20568. https://doi.org/10.1007/s00521-023-08863-9

Hendry, D., Chai, K., Campbell, A., Hopper, L., O'Sullivan,

P., & Straker, L. (2020). Development of a Human

Activity Recognition System for Ballet Tasks. Sports

medicine - open, 6(1), 10. https://doi.org/10.1186/

s40798-020-0237-5

Hölzemann, A., & Van Laerhoven, K. (2018). Using Wrist-

Worn Activity Recognition for Basketball Game

Analysis. In iWOAR '18: Proceedings of the 5th

International Workshop on Sensor-based Activity

Recognition and Interaction, 1-6. https://doi.org/10.

1145/3266157.3266217

Hsu, Y., Yang, S., Chang, H., & Lai, H. (2018). Human

Daily and Sport Activity Recognition Using a Wearable

Inertial Sensor Network. IEEE Access, 6, 31715-31728.

https://doi.org/10.1109/ACCESS.2018.2839766

Keane, J., Malone, S., Keogh, C., Young, D., Coratella, G.,

& Collins, K. A. (2021). Comparison of

Anthropometric and Performance Profiles between

Elite and Sub-Elite Hurling Players. Applied Sciences,

11(3):954. https://doi.org/10.3390/app11030954

Kranzinger, C., Bernhart, S., Kremser, W., Venek, V.,

Rieser, H., Mayr, S., & Kranzinger, S. (2023).

Classification of Human Motion Data Based on Inertial

Measurement Units in Sports: A Scoping

Review. Applied Sciences, 13, 8684. https://doi.org/10.

3390/app13158684

Kulsoom, F., Narejo, S., Mehmood, Z., Mehmood, Z.,

Chaudhry, H., Butt, A., & Bashir, A. (2022). A review

of machine learning-based human activity recognition

for diverse applications. Neural Computing &

Applications, 34, 18289–18324. https://doi.org/10.

1007/s00521-022-07665-9

Leddy, C., Bolger, R., Byrne, P. J., Kinsella, S., & Zambrano,

L. (2023). Concurrent validity of the human pose

estimation model "MediaPipe Pose" and the XSENS

inertial measuring system for knee flexion and extension

analysis during hurling sport motion. In 2023 IEEE

International Workshop on Sport, Technology and

Research (STAR) (pp. 49-52). Cavalese - Trento, Italy.

https://doi.org/10.1109/STAR58331.2023.10302442

Liu, Z., Kong, J., Qu, M., Zhao, G., & Zhang, C. (2022).

Progress in Data Acquisition of Wearable

Sensors. Biosensors, 12(10), 889. https://doi.org/10.

3390/bios12100889

McDevitt, S., Hernandez, H., Hicks, J., Lowell, R.,

Bentahaikt, H., Burch, R., Ball, J., Chander, H., Freeman,

C., Taylor, C., & Anderson, B. (2022). Wearables for

Biomechanical Performance Optimization and Risk

Assessment in Industrial and Sports

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation

Applications. Bioengineering (Basel, Switzerland), 9(1),

33. https://doi.org/10.3390/bioengineering9010033

McGrath, J., Neville, J., Stewart, T., Clinning H. & Cronin,

J. (2021). Can an inertial measurement unit (IMU) in

combination with machine learning measure fast

bowling speed and perceived intensity in cricket?

Journal of Sports Sciences, 39:12, 1402-1409.

https://doi.org/10.1080/02640414.2021.1876312

Mohsen, S., Elkaseer, A., & Scholz, S. G. (2022). Human

activity recognition using K-nearest neighbor machine

learning algorithm. In S. G. Scholz, R. J. Howlett, & R.

Setchi (Eds.), Sustainable design and manufacturing.

KES-SDM 2021. Smart innovation, systems and

technologies (Vol. 262). Springer. https://doi.org/10.

1007/978-981-16-6128-0_29

Movella, (17

January 2022). Sensor-Placement-in-Xsens-

Link-System. Movella.com. Retrieved August 27,

2024, from https://base.movella.com/s/article/Sensor-

Placement-in-Xsens-Link-System?language=en_US

Movella (2024). XSENS MVN Link. Retrieved May 26,

2024, from https://www.movella.com/products/motion

-capture/xsens-mvn-link

Muazu Musa, R., P P Abdul Majeed, A., Taha, Z., Chang,

S. W., Ab Nasir, A. F., & Abdullah, M. R. (2019). A

machine learning approach of predicting high potential

archers by means of physical fitness indicators. PloS

one, 14(1), e0209638. https://doi.org/10.1371/journal.

pone.0209638

Mullane, M., Turner, A.N., & Bishop, C. (2018). Strength

and Conditioning Considerations for Hurling: An

Amateur Gaelic Games Sport. Strength and

Conditioning Journal, 40, 72–84. https://doi.org/10.

1519/SSC.0000000000000381

Myer, G. D., Lloyd, R. S., Brent, J. L., & Faigenbaum, A.

D. (2013). How Young is "Too Young" to Start

Training?. ACSM's health & fitness journal, 17(5), 14–

23. https://doi.org/10.1249/FIT.0b013e3182a06c59

Nurwulan, N & Selamaj, G. (2020). Random Forest for

Human Daily Activity Recognition. Journal of Physics

Conference Series. 1655. 012087.

https://doi.org/10.1088/1742-6596/1655/1/012087

Nguyen, B., Coelho, Y., Bastos, T., & Krishnan, S. (2021).

Trends in human activity recognition with focus on

machine learning and power requirements. Machine

Learning with Applications, 5, 100072.

https://doi.org/10.1016/j.mlwa.2021.100072

Pajak, G., Krutz, P., Patalas-Maliszewska, J., Rehm, M.,

Pajak, I., & Dix, M. (2022). An approach to sport

activities recognition based on an inertial sensor and

deep learning. Sensors and Actuators A: Physical, 345,

113773. https://doi.org/10.1016/j.sna.2022.113773

Pfeiffer, K. A., Lisee, C., Westgate, B. S., Kalfsbeek, C.,

Kuenze, C., Bell, D., Montoye, A. H. K. (2023). Using

Accelerometers to Detect Activity Type in a Sport

Setting: Challenges with Using Multiple Types of

Conventional Machine Learning Approaches.

Measurement in Physical Education and Exercise

Science, 27(1), 60–72. https://doi.org/10.1080/

1091367X.2022.2069467

Preatoni, E., Nodari, S., & Lopomo, N. F. (2020).

Supervised Machine Learning Applied to Wearable

Sensor Data Can Accurately Classify Functional

Fitness Exercises Within a Continuous

Workout. Frontiers in bioengineering and

biotechnology, 8, 664.

https://doi.org/10.3389/fbioe.2020.00664

Ren, X., Ding, W., Crouter, S.E., Mu, Y., & Xie, R. (2016).

Activity recognition and intensity estimation in youth

from accelerometer data aided by machine

learning. Applied Intelligence, 45, 512-529.

Ribeiro, J. N., Gonçalves, B., Coutinho, D., Brito, J.,

Sampaio, J., & Travassos, B. (2020). Activity Profile

and Physical Performance of Match Play in Elite Futsal

Players. Frontiers in psychology, 11, 1709.

https://doi.org/10.3389/fpsyg.2020.01709

Rosati, S., Balestra, G., & Knaflitz, M. (2018). Comparison

of Different Sets of Features for Human Activity

Recognition by Wearable Sensors. Sensors (Basel,

Switzerland), 18(12), 4189.

https://doi.org/10.3390/s18124189

Roslan, M., & Ahmad, A. (2020). Wearable Devices

Components for High Jump Sport: An Overview.

International Journal of Integrated Engineering, 12, 27-

39. https://doi.org/10.30880/ijie.2020.12.01.003

Seçkin, A.Ç., Ateş, B., & Seçkin, M. (2023). Review on

Wearable Technology in Sports: Concepts, Challenges

and Opportunities. Applied Sciences, 13(18):10399.

https://doi.org/10.3390/app131810399

Shahar, N., Ghazali, N. F., As'ari, M. A., & Swee, T.

T. (2020). Wearable inertial sensor for human activity

recognition in field hockey: Influence of sensor

combination and sensor location. In: 2nd Joint

International Conference on Emerging Computing

Technology and Sports, JICETS 2019, 25 - 27

November 2019, Bandung, Indonesia.

https://doi.org/10.1088/1742-6596/1529/2/022015

Shouran, M., & Elgamli, E. (2020). Design and

Implementation of Butterworth Filter. International

Journal of Innovative Research in Science Engineering

and Technology, 9, 7975.

Slim, S. O., Atia, A., Elfattah, M. M. A., & Mostafa, M.-S.

M. (2019). Survey on human activity recognition based

on acceleration data. International Journal of Advanced

Computer Science and Applications, 10(3).

https://doi.org/10.14569/IJACSA.2019.0100311

Steels, T., Van Herbruggen, B., Fontaine, J., De Pessemier,

T., Plets, D., & De Poorter, E. (2020). Badminton

Activity Recognition Using Accelerometer

Data. Sensors; 20(17), 4685. https://doi.org/10.3390/s

20174685

Tran, A., Guan, J., Pilantanakitti, T., & Cohen, P.R. (2014).

Action Recognition in the Frequency Domain. ArXiv,

abs/1409.0908.

Twomey, N., Diethe, T., Fafoutis, X., Elsts, A.,

McConville, R., Flach, P., & Craddock, I. (2018). A

Comprehensive Study of Activity Recognition Using

Accelerometers. Informatics, 5(2):27. https://doi.org/

10.3390/informatics5020027

icSPORTS 2024 - 12th International Conference on Sport Sciences Research and Technology Support

Vallance, E., Sutton-Charani, N., Imoussaten, A.,

Montmain, J., & Perrey, S. (2020). Combining internal-

and external-training-loads to predict non-contact

injuries in soccer. Applied Sciences, 10(15), 5261.

https://doi.org/10.3390/app10155261

Ward, J.A., Lukowicz, P., & Gellersen, H. (2011).

Performance metrics for activity recognition. ACM

Trans. Intell. Syst. Technol., 2, 6:1-6:23.

https://doi.org/10.1145/1889681.1889687

Wolpert, D. H., & Macready, W.G. (1997). No free lunch

theorems for optimization. IEEE Transactions on

Evolutionary Computation, 1(1), 67-82. https://doi.org/

10.1109/4235.585893

Xia, C., & Sugiura, Y. (2021). Optimizing Sensor Position

with Virtual Sensors in Human Activity Recognition

System Design. Sensors. 21(20):6893. https://doi.org/

10.3390/s21206893

Yin, C., Chen, J., Miao, X., Jiang, H., & Chen, D. (2021).

Device-Free Human Activity Recognition with Low-

Resolution Infrared Array Sensor Using Long Short-

Term Memory Neural Network. Sensors, 21(10), 3551.

https://doi.org/10.3390/s21103551.

Evaluating Sensor Placement and Feature Importance for Hurling Movement Classiﬁcation