Development of a Simple Tracking System to Monitor Curling Stone

Dynamics

Shimpei Aihara

, Guanyu Chen

, Riku Hara

, Minori Nakagawa

, Ayumi Ogasawara

Hitoshi Yanagi

4,5

, Yoshinari Takegawa

, Takeshi Ito

and Fumito Masui

1,6 e

Department of Sport Science and Research, Japan Institute of Sport Sciences, Tokyo, Japan

School of Systems Information Science, Future University Hakodate, Hokkaido, Japan

School of Information Science and Technology, The University of Electro-Communications, Tokyo, Japan

High Performance Committee, Japan Curling Association, Tokyo, Japan

Common Education Group, Kitami Institute of Technology, Hokkaido, Japan

Information Communication Group, Kitami Institute of Technology, Hokkaido, Japan

Keywords: Tracking System, Deep Learning, Curling, Stone, Camera.

Abstract: Curling, frequently called chess on ice, is a sport renowned for its strategic depth, making it essential to

analyze stone dynamics for advanced tactical purposes. Systems have been developed to meticulously

measure stone position and dynamics using camera imagery and devices affixed to the rocks. However,

systems relying on multiple dedicated cameras and mounted devices encounter challenges related to

portability and simplicity. Therefore, this paper presents the development of a simple tracking system

employing a single camera to measure stone position and dynamics. The system uses deep learning for stone

detection and tracking and calibration for coordinate calculation to visualize stone dynamics. This paper aims

to support training and tactical planning using this tracking system. Experiments were conducted using the

tracking system to assess its accuracy for static stones and the velocity of moving stones. The results indicated

an average positional accuracy error of 0.02 m and an average velocity accuracy of 0.05 m/s, demonstrating

the tracking system’s high accuracy and practical feasibility.

1 INTRODUCTION

Curling is a sport where stones are thrown toward a

target, called the house, on an icy surface, competing

to score points. This sport was officially introduced

as a medal event at the 1998 Nagano Winter

Olympics. Since then, it has gained popularity

worldwide. The Japanese women’s curling team

made history by winning Japan’s first Olympic

bronze medal in curling at the 2018 PyeongChang

Winter Olympics. They followed it with a silver

medal at the 2022 Beijing Winter Olympics,

showcasing their talent on the global stage. As they

aim for further glory, there are high expectations for

improvements in their competitive abilities.

https://orcid.org/0000-0002-8513-0204

https://orcid.org/0009-0007-8754-6979

https://orcid.org/0000-0003-1947-0021

https://orcid.org/0009-0001-5253-1036

https://orcid.org/0000-0001-9979-8734

Curling, a sport renowned for its strategic and

technical challenges, tests the precision and tactics of

its players. The accuracy of each shot and the tactics

employed can significantly impact the outcome of a

match, directly affecting the team’s victory or defeat.

In matches at the level of the Japanese national team,

a significant correlation was observed between the

accuracy of shots and the points scored. Specifically,

shot accuracy influences the point difference in

matches (Masui, 2016), underscoring the importance

of precision in shots to execute tactics tailored to the

unfolding dynamics of the game. Thus, understanding

the behavior of stones, which dictates the accuracy of

shots, is paramount for executing precise shots.

Aihara, S., Chen, G., Hara, R., Nakagawa, M., Ogasawara, A., Yanagi, H., Takegawa, Y., Ito, T. and Masui, F.

Development of a Simple Tracking System to Monitor Curling Stone Dynamics.

DOI: 10.5220/0012896100003828

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 22-33

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

Various systems using camera video files have

been proposed to measure stone position and behavior.

For instance, at the 2022 Beijing Winter Olympics

and Paralympics, a system using video files from 42

dedicated cameras installed at the competition venue

was introduced to measure stone position and

trajectory (Shi, 2022). However, such systems

requiring multiple cameras and dedicated installation

at the competition venue pose challenges for

simplicity and portability. Additionally, their use is

limited to specific curling halls, further complicating

their portability.

Methods have also been implemented by

attaching devices called inertial measurement units

(IMUs) to the handles of curling stones to measure

their velocity, angular velocity, and displacement

(Lozowski, 2016). However, attaching devices to

curling stones might alter their original weight,

potentially affecting their behavior. Moreover, such

attachments are prohibited in official competitions,

rendering them ineffective for measuring the

behavior of stones used in tournaments.

Given the challenges of existing systems, an

urgent need exists to introduce markerless, portable,

and easy-to-use tracking systems in the competitive

arena. Therefore, we propose a simple tracking

system that uses highly portable cameras to provide

feedback on the behavior of curling stones. This

system aims to facilitate the easy measurement of

stone behavior, contributing to training and tactical

support.

This report discusses the accuracy evaluation of

the proposed simple tracking system’s analysis of

stone velocity compared to laser velocity

measurement equipment. The results will illuminate

the proposed system’s usefulness and accuracy,

guiding further research and practical implementation

efforts.

2 PROPOSED METHOD

The proposed simple tracking system relies on

foundational technologies, including a detection

model to identify stones within the image, a tracking

model to follow the same stone across frames, and a

calibration model to transform stone positions from

the camera coordinate system to the global coordinate

system. In this chapter, we discuss each technology

constituting the proposed system.

2.1 Stone Detection Model

The stone detection model in this study was

developed using the highly efficient You Only Look

Once (YOLO) detector as the underlying model

(Wang, 2023). YOLO is a popular object detection

algorithm known for its real-time performance and

exceptional accuracy in identifying objects within

images. Transfer learning was employed with a

custom dataset of curling stone images to fine-tune

the YOLO detector for detecting curling stones. This

approach leverages pretrained weights from a general

object detection model and adapts them to the target

domain, resulting in an optimized stone detection

model for curling scenarios.

2.1.1 Dataset Creation

We recorded videos of competitions and practice

sessions at various curling halls for dataset creation.

An example of the filming setup is illustrated in

Figure 1. We used commercially available video

cameras, smartphones, and tablets to film from

different vantage points near the ice sheet and

spectator areas, varying the camera placement,

angles, and shooting conditions, such as aperture,

shutter velocity, sensitivity, and white balance, to

ensure diversity in the acquired data. The video file

encompassed various environments, including

variations in the design and logos of the curling sheet

houses, the colors and patterns of the background

walls, and the lighting conditions. Additionally, since

some curling rinks have glass windows separating the

playing area from the spectator area, we included

video files shot through glass.

Subsequently, we randomly extracted images

from the captured video files and annotated the stones

within the images. We surrounded the areas of the

stones, excluding the handle parts, with bounding

boxes and labeled them as red stone or yellow stone.

Figure 2 shows an example of images from the

dataset. We ensured coverage of various scenarios

expected in curling scenes, including images where

players overlapped with stones, brushes overlapped

with stones, and stones overlapped with each other.

Staff trained in annotation tasks performed the

dataset creation. Furthermore, multiple staff members

mutually reviewed the annotated data, creating a

high-quality dataset. This dataset, which is

proprietary to us, is a collection of data and a crucial

tool for developing a detection model.

Development of a Simple Tracking System to Monitor Curling Stone Dynamics

Figure 1: Scene of the taking videos.

Figure 2: Sample images of the curling stone dataset.

2.1.2 Training the Detection Model

The YOLO v7 framework (Wang, 2023), known for

its broad applicability and flexibility in object

detection tasks, was the foundational model for our

detection model. We initialized the model with

parameter weights obtained from pretraining on the

Microsoft Common Objects in Context (Lin, 2014), a

comprehensive dataset of red–green–blue (RGB)

images. Then, fine-tuning was conducted using our

proprietary dataset comprising curling stone images.

We used the Adam optimization algorithm

(Kingma, 2015) with a learning rate of 0.01 to

optimize the model parameters. Thirty percent of the

dataset was reserved for validation, and we selected

the model parameters that yielded the lowest loss on

the validation set. The training was performed on a

PC (ELSA GALUDA G5-ND G450E) equipped with

a Core i7-13700K CPU, GeForce RTX 4090 GPU,

and 64 GB of memory. The programming language

was Python, and the machine learning library was

PyTorch.

The model was designed to take an RGB image as

input and output the stone coordinates in the image

and their respective colors (red and yellow).

2.2 Stone Tracking Model

The detection model identifying the stones in each

image alone sometimes fails to detect stones due to

occlusion caused by overlapping players or other

stones and motion blur resulting from stone

movement. Therefore, we implemented a model that

tracks the same stones across video frames and fills

in the missing frames by correlating consecutive

detection results. In this study, we developed our

model based on ByteTrack (Zhang, 2022), a method

for tracking multiple objects. ByteTrack achieves

stability, velocity, and accuracy through its Kalman

filter–based algorithm. However, it must improve

accuracy during complex movements and needs

appearance information to prevent each stone switch

identifier from switching. Therefore, we combined

the motion and visual information of ByteTrack using

attention mechanisms to achieve highly accurate

tracking. This step enables the stable tracking of

individual stones in the unique curling environment,

where stones might become obscured by sweeping or

where they might collide, resulting in complex stone

movements.

2.3 Calibration Model

We developed a calibration model to transform the

stone coordinates in the images to their global

positions on the ice sheet. Obtaining coordinate

values for calibration reference points in a real game

or practice venue frequently requires more time. Also,

calibration tools, such as poles or checkerboards,

which are common in lab settings, could be more

practical.

Figure 3: Screen capture of the calibration tool.

Therefore, we devised a method to calculate

camera parameters from distinctive points visible in

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

the venue captured by the camera. In the game venues

or arenas, lines are drawn on the court according to

the rules regarding length and size, such as the circular

mark called the house and the line known as the hog

line in curling. Using known ice sheet information, we

created a calibration tool (ice sheet model) to calculate

camera parameters by providingthe corresponding

global coordinates for pixel coordinates (at least four

points) in the RGB images. Figure 3 shows a

screenshot of the developed calibration tool.

We used the Levenberg–Marquardt method

(Moré, 1977) to calculate the camera’s position and

orientation in three-dimensional space and the hill

climbing method (Goldfeld, 1966) to determine the

camera’s focal length, representing the internal

camera parameters. With these camera parameters

and the predefined three-dimensional shape

information of the stone, we calculated the

coordinates of the midpoint of the stone’s bottom

surface on the ice sheet.

The calibration model does not require the entire

curling sheet to be visible in the video file.

Calibration can be performed as long as a portion of

the curling sheet, such as a part of the house or the

lines, is visible, offering the significant advantage of

using video files filmed under any conditions. For

example, even if the video file is a close-up of a

specific area, analysis can be conducted if part of the

curling sheet is visible. Therefore, video files

captured from various camera positions, such as the

audience seats or the coach’s area, can be

accommodated. Furthermore, the calibration process

allows accurate position data to be obtained, even if

the filming angle varies, enabling the precise analysis

of video files taken from oblique or irregular angles.

Our system is robust and reliable. It can analyze

video files where only part of the house or lines are

visible. This advantage makes it usable even if other

objects are in the frame or part of the view is

obstructed, ensuring that accurate data can be

obtained in real-world match or practice

environments where obstacles or other players’

movements might affect the video file.

This system’s flexibility is crucial for curling

analysis. For example, it can consistently analyze

video files from different venues, matches, or practice

sessions with various camera settings and conditions,

ensuring data consistency and reliability.

3 EVALUATIONS

We developed a tracking method for curling stones

using the detection model described in Section 2.1,

the tracking model described in Section 2.2, and the

calibration model described in Section 2.3. Our

method processes video footage a monocular camera

captures to detect and track the curling stone,

enabling us to calculate the stone’s position in sheet

coordinates.

Three evaluations were conducted to validate the

accuracy of the proposed method. These evaluations

included the accuracy validation of the stone’s static

position, the split time when the stone passed through

specified intervals, and the stone’s velocity. Section

3.1 discusses the accuracy validation of the stone’s

position, Section 3.2 covers the accuracy validation

of the split time, and Section 3.3 explains the

accuracy validation of the stone’s velocity.

3.1 Evaluation of Stone Positions

3.1.1 Data Acquisition Experiment

The experiment was conducted at the Argo Graphics

Kitami Curling Hall. The proposed method was used

to detect the positions of curling stones placed in

known absolute coordinates with high accuracy. First,

eight stones were placed on the ice sheet (Figure 4).

The stones are labeled from A to H. The coordinate is

centered at the middle of the ice sheet, with units in

meters.

The recorded video file was analyzed using the

proposed method to detect the stone positions. The

positions detected were then compared to the actual

positions of the static stones to evaluate the accuracy.

Figure 4: Position and coordinates of stones to be placed.

Development of a Simple Tracking System to Monitor Curling Stone Dynamics

Figure 5 shows a frame from the recorded video file

used for the evaluation. The accuracy evaluation

involved calculating the differences between the

static stone actual position data and the positions

calculated using the proposed method. These

positional discrepancies were statistically analyzed to

determine the method’s accuracy.

Figure 5: Image of stones placed to evaluate positional

accuracy.

3.1.2 Results

The accuracy evaluation results of the static positions

of stones calculated using the proposed method are

discussed. Table 1 shows the correct coordinates for

eight stones, the coordinates calculated using the

proposed method, and the error values. The “Correct

Position” refers to the known 2D coordinates of

points A–H on the curling sheet (Figure 4). The

“Calculated Position” indicates the 2D coordinates

determined using the proposed method. The “Error”

represents the discrepancy between these two values.

The L2 norm in the Error column indicates the

Euclidean distance between the correct and calculated

coordinates. Consequently, the average tracking

system error for the eight stones was 0.02 m. The

minimum error was 0.00 m for stone D, located

approximately 7 m from the camera, and the

maximum error was 0.03 m for stones G and H.

3.1.3 Discussions

The system used to measure the position and behavior

of curling stones at the 2022 Beijing Olympics and

Paralympics reported an accuracy of 0.30 ± 0.03 m

over a measurement range exceeding 20 m (Shi,

2022). However, our proposed method achieves

higher precision than previous technology, providing

practical accuracy for real-world applications.

Specifically, our method calculates the behavior of

stones using video footage captured using a

monocular camera, significantly improving the ease

of installation and operation compared to a previous

study (Shi, 2022), which required the permanent

installation of multiple specialized cameras at the

venue.

Furthermore, our proposed method can detect the

position of a stone approximately 40 m away from the

camera with an error margin of only 0.03 m. This high

accuracy level is attributed to the high-quality dataset,

as described in Section 2.2.1, where even the most

petite stones in the images (those with few pixels)

were annotated. This capability to detect stones

across the entire sheet using only a single monocular

camera has not been reported previously and is a

highly innovative approach.

Traditionally, players and coaches recorded the

stone positions manually using notebooks, which was

time-consuming and limited in accuracy. Typically,

positions were recorded with a granularity of

approximately half to one stone’s width, leading to

significant errors due to visual estimation. The

proposed system can now record the stone positions

simply by capturing video footage, resulting in

simplicity and accuracy. This method provides much

more precise data than manual recording, leading to

significant advancements in analyzing curling

strategies and techniques. Moreover, this method

allows players and coaches to analyze practice and

game performances more efficiently. The automated

system reduces human errors and saves time and

Table 1: Errors between the correct values of the stone position and the calculated values of the stone tracking system.

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

Figure 6: Image of evaluation indexes.

effort, enabling players to focus on improving their

techniques, refining their strategies, and enhancing

their overall performance.

3.2 Evaluation of Split Times

3.2.1 Data Acquisition Experiment

In curling, the hog–hog split time (H–H time) is

typically used for evaluating stone velocity. As

shown in Figure 6, the H–H time refers to the time a

curling stone travels between the hog lines (a distance

of 21.94 m) marked on the ice surface. Understanding

the correct velocity is crucial for ensuring that the

stone reaches its intended target accurately. During a

match, the H–H time can be used to analyze players’

performance in detail and make tactical adjustments.

For example, players can adjust their throws based on

different ice conditions and determine the optimal

velocity for specific strategies.

In practice, accurately measuring the stone’s

velocity helps players improve the consistency of

their throws, which is vital for skill development and

increasing performance. Therefore, measuring the H–

H time is crucial in curling. During training, the H–H

time is measured using laser velocity measurement

equipment (Rock Hawk – 2 beam), which is widely

used in actual practice environments and has become

a de facto standard used by most teams worldwide.

The players and coaches use stopwatches to measure

the H–H time during matches.

For the data acquisition experiment to evaluate

velocity accuracy, we recorded videos of the stone

being thrown and simultaneously measured the H–H

time using the laser velocity measurement equipment.

We then compared the H–H time calculated using the

proposed method with the H–H time measured by the

laser velocity measurement equipment. This compare-

son allowed us to assess the accuracy of the method.

During the experiment, shots with varying

velocities were measured to ensure the collected data

were balanced and bias-free. Specifically, stones

were deliberately thrown at different velocities to

cover various velocities. Additionally, yellow and red

stones were used to prevent color–based bias in the

data. We collected data from 66 shots.

We used an iPhone 15 pro max to record the

videos, positioning it approximately 1.4 m above the

edge of the sheet (Figure 7 (left)). The resolution was

set to 4 K, and the frame rate was 60 frames per

second (fps). Figure 7 (right) shows the image from

the recorded video. The laser velocity measurement

equipment used was the same as in actual practice

settings (Rock Hawk – 2 beam). Figure 7 (right)

shows that the laser velocity measurement equipment

was set up at the hog line, and the H–H time was

displayed and recorded using an application

developed by Tim Senger.

3.2.2 Results

In this evaluation, we compared the H–H time

calculated using the proposed method (vertical axis)

with that obtained using a laser velocimeter, consi-

dered the ground truth (horizontal axis) (Figure 8).

The results indicate a remarkably high correlation

coefficient of 1.00 between the H–H times measured

using the proposed method and the laser velocimeter.

This result demonstrates that the proposed system

achieves highly accurate H–H time measurements.

Specifically, the mean absolute error between the H–

H times calculated using the proposed method and the

ground truth measured by the laser velocimeter was

0.10 seconds, with a standard deviation of 0.07

seconds.

Figure 7: Camera setting (left). The sample image from the

recorded video (right).

Development of a Simple Tracking System to Monitor Curling Stone Dynamics

Figure 8: The H–H time correlation between the proposed

method and the laser velocimeter.

3.2.3 Discussions

These results indicate that the H–H time measurement,

which traditionally relied on expensive specialized

equipment, such as laser velocimeters, can now be

effectively achieved using only video recordings.

While laser velocimeters offer precise velocity

measurements, they are costly and require significant

effort to set up. However, the proposed system uses

video, making the measurement process more

convenient and reducing economic burdens.

Furthermore, the proposed method is beneficial in

actual competitive environments. Athletes and

coaches frequently resort to manual timing with

stopwatches, influenced by the measurer’s

proficiency and fatigue and whether they use their

dominant or nondominant hand. A study evaluating

the errors in manual timing with stopwatches for 50-

m sprints reported a systematic error of 0.27 s (Todou,

2018). Although the sports differ, this result indicates

that our proposed model achieves higher precision

than manual timing. The proposed method enables

the swift and accurate measurement of the H–H time,

which is crucial for evaluating player performance

and conducting strategic analyses during

competitions. This capability facilitates faster tactical

decision-making and provides valuable feedback for

enhancing player skills.

The proposed method has been confirmed as a

viable alternative to traditional laser velocimeters,

offering high-precision H–H time measurements.

This advancement can significantly improve

performance analysis methods in curling, leading to

more efficient and effective training and strategy

development.

3.3 Evaluation of Velocity

3.3.1 Data Acquisition Experiment

The velocity at which a stone crosses the hog line is

typically used in curling. The hog line crossing

velocity refers to the stone’s velocity as it passes over

the hog line marked on the ice surface of the curling

rink (Figure 7). Players must release the stone before

it crosses the hog line, so this velocity is the

approximate initial velocity of the stone in curling.

The initial stone velocity is crucial to ensure it

reaches the target accurately. Throwing the stone at

an appropriate velocity increases the likelihood of it

reaching the desired destination. If the velocity is too

high, the stone might overshoot the target, whereas

the stone might not reach it if it is too low. Therefore,

it is essential to measure and adjust the stone’s

velocity accurately.

Accurately managing the throwing velocity

enables players to achieve consistent throws. This

consistency helps predict the stone’s trajectory and

allows for strategic play. For example, the overall

team tactics can be effectively executed by

determining the optimal velocity for specific

strategies. The hog line crossing velocity is measured

using a laser velocimeter during training. This device,

used in competitive environments, provides high-

precision velocity measurements. The laser

velocimeter is critical for supporting players’ skill

development, as it allows for the precise

measurement of their throwing velocities.

The data collection experiment for velocity

evaluation was conducted as follows. First, videos of

the thrown stones and the corresponding hog line

crossing velocities were recorded using a laser

velocimeter. Subsequently, the hog line crossing

velocities calculated using the proposed method were

compared with those measured by the laser

velocimeter.

This experiment collected 66 shot data points.

Figure 7 shows that the camera and laser velocimeter

measurement conditions were consistent with those

described in Section 3.2.1, ensuring uniformity in the

data collection.

3.3.2 Results

The evaluation of the hog line crossing velocities

yielded insightful results. A comparative analysis was

conducted between the velocities calculated using the

proposed method and those obtained through a laser

velocimeter. Figure 9 illustrates this comparison

through a correlation plot, where the vertical axis

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

denotes the hog line crossing velocities computed

using the proposed method, and the horizontal axis

represents the ground truth velocities measured by the

laser velocimeter.

The evaluation revealed a high correlation

coefficient of 0.98 between the velocities derived

from the proposed method and those measured by the

laser velocimeter, indicating exceptional accuracy in

velocity measurement. Specifically, the absolute

mean error between the hog line crossing velocities

calculated using the proposed method and the ground

truth velocities measured by the laser velocimeter was

0.05 m/s, with a standard deviation of 0.05 m/s.

Figure 9: Correlation of the hog line crossing velocities

between the proposed method and the laser velocimeter.

3.3.3 Discussions

These results demonstrate the practical effectiveness

of the proposed method, particularly in accurately

estimating the speed of stones more than 37 m away

from the camera. Traditionally, measuring the speed

of a stone crossing the hog line required expensive

specialized equipment, such as laser velocimeters.

However, this study shows that such measurements

can be effectively performed using only video

recordings, meaning a significant reduction in cost

and setup complexity compared to traditional

methods.

Furthermore, the accuracy of the proposed

method improves as the distance between the camera

and the stone decreases. The closer the camera is to

the stone, the higher the resolution of the video,

allowing for more detailed data collection. Figure 10

shows an example of a video image zoomed in on the

delivery motion, which is the action of throwing the

stone. Figure 11 shows the analysis results using the

proposed method. The correlation with the actual

value measured by the laser velocimeter improved to

1.00. The mean absolute error improved to 0.02 m/s,

and the standard deviation improved to 0.01 m/s.

Thus, when conducting detailed delivery analysis,

changing the shooting conditions to capture larger

images of the stone can yield more accurate tracking

data, allowing for objectively evaluating slight

fluctuations and accelerations or decelerations during

delivery.

Moreover, while traditional methods using laser

velocimeters were limited to fixed–point speed

measurements, the proposed method can

continuously record the speed of the stone from the

start of the shot until it stops, enabling a

comprehensive analysis of the entire shot.

This advancement is highly significant for curling

training and competition. Coaches and players can

now analyze the movement of the stones in greater

detail, enabling more precise technical guidance and

strategic planning. By continuously recording the

speed data of the stones, subtle adjustments can be

made to the players’ techniques and team strategies,

improving overall performance.

Additionally, the proposed method allows for a

detailed analysis of the relationship between the

stone’s speed and curl, which is critical for gaining

new insights into curling strategy and technique. For

example, it becomes possible to analyze how the

stone’s curl changes at different speeds and identify

the optimal shots within specific speed ranges to

achieve the desired curl.

Therefore, the proposed method opens a new

dimension in data analysis for curling, providing

coaches and players with a powerful tool for more

efficient and effective training and strategy

development. In addition, this method should

improve individual player skills and enhance the

strategic play of teams, raising the overall

competitive level of curling.

Figure 10: Example of the zoomed-in video image of the

delivery operation.

Development of a Simple Tracking System to Monitor Curling Stone Dynamics

Figure 11: Correlation of the hog line crossing velocities

between the proposed method and the laser velocimeter.

The case of using a zooming video.

4 IMPLEMENTATIONS OF THE

TRACKING SYSTEM

4.1 Overview of the Curling Stone

Tracking System

We developed a tracking system for curling stones

using the detection model described in Section 2.1,

the tracking model described in Section 2.2, and the

calibration model described in Section 2.3. Figure 12

shows the overview of the proposed system. This

system uses the Swift programming language and

operates on the Mac operating system (macOS). This

system reads a video file captured by a monocular

camera, tracks curling stones within the video, and

visualizes the tracking results, allowing users to

efficiently operate the system on a single computer

and analyze the movements of curling stones.

Here is an explanation of the usage flow of this

tracking system. First, the users load the video file

they wish to analyze into the system. Next, the

calibration model calculates the camera parameters

from the video image. This calibration converts the

coordinates in the video image to actual coordinates

on the ice sheet. Then, using the stone detection and

stone tracking models, the system detects and tracks

the curling stones in the video, allowing the system to

obtain time-series data on the positions of the curling

stones on the ice sheet. Also, the time-series position

data, velocity data, stopping position data, and other

relevant information for each stone can be output as

Figure 12: Overview of the tracking system.

Comma Separated Values (CSV) files for data storage

and later analysis.

The processing velocity of the system achieves

over 30 fps for 4 K video files, which was realized on

a MacBook Pro (specifications: 10-core CPU, 16-

core GPU, 32 GB memory). This high-velocity

processing enables faster tracking of stone

movements.

The system can analyze.mp4 and.mov format files

captured by smartphones, tablets, video cameras, and

other devices, allowing users to use video files

captured on various devices. The system supports

various filming conditions, such as the shooting

position (e.g., spectator seats, coach seats, beside the

ice sheet), the orientation of the ice sheet (front, side,

diagonal), and whether the video file is shot through

glass, ensuring accurate analysis from different

viewpoints and angles. Additionally, the system can

simultaneously analyze multiple stones on the ice

sheet, allowing the tracking and analysis of the

movements of multiple stones during matches.

Moreover, the system can be used for matches and

practice sessions to analyze curling scenes in any

context. Although curling venues might have different

designs for the house on the ice sheet and various

sponsor logos, the system can analyze video files from

any curling rink. This advantage ensures consistent

analysis results in different facilities and conditions.

4.2 Visualization of Curling Stone

Dynamics

We visualized the stone dynamics from the obtained

time-series position data. Figure 13 shows an

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

Figure 13: Examples of visualization.

example of stone dynamics visualization drawn using

this tracking system. Additionally, it can draw

overhead views of the stone movement trajectories.

Figure 13 shows that the system can calculate and list

each stone’s stopping position and the velocity at

which it crosses specified lines.

Thus, this tracking system is a powerful tool for

detailed analysis and the clear visualization of curling

stone movements during matches and practice

sessions. It is an innovative system that has not been

reported before. This system let users quickly obtain

data on stone movements, velocities, stopping

positions, and more, aiding in game analysis and

strategy planning. For example, by reviewing stone

movements after a match, players can identify areas

for improvement and measure the effectiveness of

their practice. Additionally, the system can provide

specific feedback to players during coaching.

Figure 14 shows that this system can analyze the

trajectory and speed of curling stones and generate

computer-generated renderings using only footage

from a single camera. This capability opens

possibilities for applications in television broadcasts

and other media. Such graphics, unprecedented in

traditional curling broadcasts, offer viewers a new

way to enjoy the sport. For example, viewers can

track the stone movements in real time, visually

comprehend their trajectories and speeds, and better

understand the strategies involved. Additionally,

commentators can provide increased detailed and

specific analyses based on this data, enhancing the

viewing experience.

Moreover, this system achieves an analysis speed

of over 30 fps for 4 K footage, operating on a

MacBook Pro (specifications: 10-core CPU, 16-core

GPU, 32 GB memory). The high processing speed

makes it suitable for large-scale tournaments and

international matches, where real-time analysis is

crucial, providing viewers with a seamless, lag-free

experience. The ability to perform advanced analyses

on standard hardware is a significant advantage.

This system is valuable for analyzing and

visualizing curling stone movements. It has many

applications, from match analysis and strategy

planning to television broadcasts. Implementing this

system promises to enrich the experience of playing

and watching curling, making it more engaging and

insightful.

Figure 14: Values of stone dynamics analysis.

5 CONCLUSIONS

In this study, we developed a tracking system

comprising a detection model to detect stones from

images, a tracking model to track the same stone

within a video, and a calibration model to convert

stone positions from camera coordinates to global

coordinates. We achieved a high level of accuracy,

with an average position error of 0.02 m. Additionally,

we demonstrated the practicality of the system by

showing a strong correlation between metrics, such as

the H–H time and hog-line crossing velocities, with

those obtained from a laser velocimeter a de facto

standard. Furthermore, we successfully visualized the

dynamics of the stones from the position data.

The proposed system offers various beneficial

aspects for curling:

・ Training and skill improvement: Players can

analyze stone behavior using the tracking

system to enhance their skills, such as

understanding stone speed and curl direction.

・ Tactical and strategic improvement: Coaches

and teams can use the tracking system to

enhance their game strategies by analyzing

stone movement and position data and studying

opponents’ strategies and play styles.

Development of a Simple Tracking System to Monitor Curling Stone Dynamics

・ Match recording and analysis: The tracking

system can record and analyze match data,

enabling players and teams to identify tactical

errors and areas for improvement for future

matches.

・ Player evaluation and selection: The system

helps objectively evaluate players’ abilities and

performances, facilitating team selection and

role assignment.

・ Entertainment for viewers: The tracking system

can provide entertaining content through

television broadcasts or online streaming,

enhancing viewers’ understanding and

enjoyment of the game.

・ Fair judging in competitions: The tracking

system accurately records stone movements,

supporting fair judging by providing crucial

information to referees and event organizers.

・ Support for event management: The system can

assist in event management by recording match

progress and results, facilitating smoother event

operations.

However, the tracking system also faces several

challenges:

・ Accuracy issues: Environmental factors, such as

camera placement and lighting conditions, can

affect the system’s accuracy, primarily when

relying on a single camera to detect stones.

・ Real-time processing constraints: The proposed

system’s current workflow involves postcapture

analysis on a laptop, limiting real-time data

processing. Enhancing real-time analysis, such

as on smartphones or tablets, would

considerably benefit the system.

・ Tracking and prediction capability: Currently,

the system works by using computer vision to

track stones, providing crucial position

information. However, the potential for utilizing

this data in more advanced analytical contexts,

such as predictive modeling and tactical

decision support, has yet to be fully explored.

With the rapid development of machine

learning technologies, such as Long Short-Term

Memory (LSTM) networks, which excel in

learning and predicting time series data, there is

significant potential to design a system that

leverages existing data to develop a tracking and

prediction system. As illustrated in Figure 15,

by utilizing the outcomes of this research along

with terminal devices, we can not only visualize

the current motion state of the stone in real-time,

displaying information such as speed and angle,

but also predict the stone’s future trajectory

based on its initial motion state, using machine

learning. This would provide athletes with more

valuable information to enable more tactical

interaction with the stone. The goal is to allow

athletes not only to receive real-time feedback

on the stone’s movement but also to quickly

understand potential motion trends and make

informed actions based on experience and

training, thereby improving performance.

Moreover, the system’s predictive capabilities

could be extended to assist coaches in

developing more precise training regimens by

analyzing historical data and identifying

patterns that correlate with successful outcomes,

thereby facilitating continuous improvement

beyond immediate match scenarios.

Figure 15: A tablet device on the stone visualizes the

current stone dynamics and future trajectory.

・ Enhancements to the user interface are

necessary to make the system’s operation and

data analysis more user-friendly and accessible.

We aim to enhance the tracking system’s

performance and use and promote its adoption in

curling competitions, training, and events by

resolving these issues.

ACKNOWLEDGEMENTS

This work was supported by the "The Enhancement

of HPSC Infrastructure through Technology

Innovation Project" of Japan Sports Agency.

REFERENCES

Goldfeld, S. M., Quandt, R. E., & Trotter, H. F. (1966).

Maximization by quadratic hill-climbing.

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

Econometrica, 34(3), pp. 541–551. https://doi.org/

10.2307/1909768

Kingma, D. P., & Ba, J. (2014). Adam: A method for

stochastic optimization. arXiv preprint

arXiv:1412.6980.

Lin, T. Y., Maire, M., Belongie, S., Hays, J., Perona, P.,

Ramanan, D., & Zitnick, C. L. (2014). Microsoft coco:

Common objects in context. In Computer Vision–

ECCV 2014: 13th European Conference, Zurich,

Switzerland, pp. 740–755.

Lozowski, E., Maw, S., Kleiner, B., Szilder, K., Shegelski,

M., Musilek, P., & Ferguson, D. (2016). Comparison of

IMU measurements of curling stone dynamics with a

numerical model. Procedia Engineering, 147, pp. 596–

601. https://doi.org/10.1016/j.proeng.2016.06.246

Masui, F., Hirata, K., Otani, H., Yanagi, H., & Ptaszynski,

M. (2016). Informatics to support tactics and strategies

in curling. International Journal of Automation

Technology, 10(2), pp. 244–252. https://doi.org/

10.20965/ijat.2016.p0244

Moré, J. J. (2006). The Levenberg-Marquardt algorithm:

implementation and theory. In Numerical Analysis.

Proceedings of the Biennial Conference Held at

Dundee pp. 105–116. https://doi.org/10.1007/BFb0067

700

Shi, X., Wang, Q., Wang, C., Wang, R., Zheng, L., Qian,

C., & Tang, W. (2022). An AI-based curling game

system for winter Olympics. Research, 2022, 9805054.

https://doi.org/10.34133/2022/9805054

Todou, D., Manabe, Y., & Arakawa, H. (2018). The

quantification of the systematic and random errors by

manual measurement with a stopwatch in 50 m sprint

test. Human Performance Measurement, 18, pp. 23–33.

Wang, C. Y., Bochkovskiy, A., & Liao, H. Y. M. (2023).

YOLOv7: Trainable bag-of-freebies sets new state-of-

the-art for real-time object detectors. In Proceedings of

the IEEE/CVF Conference on Computer Vision and

Pattern Recognition pp. 7464–7475. https://doi.org/

10.1109/CVPR52729.2023.00721

Zhang, Y., Sun, P., Jiang, Y., Yu, D., Weng, F., Yuan, Z.,

& Wang, X. (2022). Bytetrack: Multi-object tracking

by associating every detection box. In European

Conference on Computer Vision, pp. 1–21.

Development of a Simple Tracking System to Monitor Curling Stone Dynamics