Development of Monocular Vision-Based Tracking Method for

Wheelchair Sports

Shimpei Aihara

, Takara Sakai

and Akira Shionoya

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

Department of Management and Information Systems Science, Nagaoka University of Technology, Niigata, Japan

Keywords: Wheelchair Sports, Positioning System, Monocular Camera, Deep Learning.

Abstract: Recently, tracking systems to measure player positions have been introduced in the sports domain. However,

wheelchair sports have not been considered extensively. In addition, user-friendly and low-cost systems for

wheelchair sports are uncommon. Thus, in this paper, we propose a method to calculate the kinematic data of

wheelchair athletes on a playing field (i.e., player positions and wheelchair directions) using images acquired

by a monocular camera. The proposed method was evaluated experimentally, and the root mean square error

of the position accuracy was 0.11 m, and the mean average error of the direction accuracy was 6.78 degrees.

The results demonstrate that the proposed method outperforms existing tracking methods in terms of accuracy.

The findings of this study suggest that it is possible to acquire kinematic data of wheelchair athletes using a

simple method, which we expect to contribute to improvement analysis of the wheelchair athlete performance.

1 INTRODUCTION

Sports promote mental and physical development,

enrich humanity, and play an important role in living

a healthy life. For people with disabilities, sports can

be an important component of their medical

rehabilitation. In addition, sports can provide lifelong

recreation and can be played at various levels

including highly competitive ones, e.g., the

Paralympics. In Japan, particularly in competitive

sports, interest in sports for the disabled increased due

to success of the Tokyo 2020 Olympic and

Paralympic Games. Wheelchair sports accounted for

about 50% of the competitions held at the Paralympic

Games, e.g., tennis, basketball, athletic sports,

badminton, rugby, and table tennis. Wheelchair sports

are recognized as an international sport, and global

competitiveness has advanced signiﬁcantly in recent

years (Perret, 2017).

Moreover, in recent years, there has been a

growing trend of utilizing technology in the field of

sports. Various technologies are used to monitor

performance in both competition and training to

realize competitive advantages (Halson, 2014). In

wheelchair sports, wheelchair movement

performance is critical to evaluate game performance

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

and optimize training routines (van der Slikke, 2016);

however, literature related to wheelchair sports is

limited compared to that of other Olympic events, and

the quantitative evaluation of wheelchair movement

performance is insufficient (Perret, 2017). Thus, the

goal of this study is to realize an affordable tracking

system to obtain kinematic data of wheelchair sports

to improve wheelchair movement performance

assessment. However, the literature related to

wheelchair sports is scarce, and the quantitative

evaluation of wheelchair movement performance is

insufficient (Perret, 2017)．This study aims to realize

a tracking system to obtain kinematic data of

wheelchair sports in order to improve the level of

movement performance assessment.

2 RELATED WORK

With the increasing competitiveness in wheelchair

sports, the utilization of technologies has expanded

(Grogan, 2012) (Laferrier, 2012). For example, to

evaluate wheelchair movement performance, Inertial

Measurement Unit (IMU) sensors attached to the

wheelchair are used frequently due to their user-

friendliness and low cost (Shepherd, 2018). IMU

Aihara, S., Sakai, T. and Shionoya, A.

Development of Monocular Vision-Based Tracking Method for Wheelchair Sports.

DOI: 10.5220/0012203600003587

In Proceedings of the 11th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2023), pages 179-186

ISBN: 978-989-758-673-6; ISSN: 2184-3201

179

sensors register movement, speed, and angular

velocity of the wheelchair player (Pansiot, 2011) (van

Dijk, 2022). In addition, methods based on IMU

sensors can measure wheelchair movement easily.

However, wheelchair positions are not accurately

defined with the help of only IMUs accurate, which

is one of the problems (van der Slikke, 2017).

The local positioning system (LPS) using wireless

technology is a method that can be used to track

wheelchair positions in various sports, e.g.,

wheelchair rugby and basketball (Rhodes, 2014), as

well as wheelchair tennis (Perrat, 2015). However, to

the best of our knowledge, very few related studies

have been reported.

Wireless LPSs measure the position and speed of a

wheelchair player at high accuracy. They comprise

many fixed base stations and mobile tags attached to

the wheelchair players; thus, there are some problems,

e.g., the need for installation of base stations at the

venue and expensive equipment. In addition, LPSs

cannot obtain wheelchair motion direction. In

wheelchair sports, the chair-work skill is an important

factor when evaluating performance; thus, wheelchair

direction information must be an available (Mason,

2013). In addition, attaching IMU sensors or LPSs to

the wheelchairs or players causes several issues, e.g.,

preventing movement. In addition, such devices are

not always permitted during official competitions.

Video-based tracking systems have been reported

in the reference. Such systems eliminate the need to

attach devices to the wheelchairs or players, and,

therefore can collect data about all athletes in the

game. In field sports, e.g., soccer, player positions can

be acquired using deep learning techniques and image

analysis from multiple cameras placed around the

field (Redwood, 2012) (Linke, 2020). These

techniques are used in FIFA and Japanese

professional league matches. However，construction

work and expensive equipment are required; thus, the

costs of such systems are high.

Therefore, less expensive methods have been

proposed to obtain player positions using single

camera images (Buric, 2019) (Zhang, 2020). In the

wheelchair rugby context, a previous study reported

the acquisition of player positions using single

camera images. Here, the wheelchair player detection

rate and the position accuracy were approximately

20% lower than in the case of soccer; thus, operator

corrections were required (Sarro, 2008). To the best

of our knowledge, no video-based tracking system

that obtains wheelchair directions has been reported

to date.

In this study, we developed a video-based tracking

method for wheelchair sports. Figure 1 shows an

outline of the proposed method. To realize a simple

and low-cost system, only a single camera is

employed in the proposed method. The proposed

system output both the wheelchair players’ positions

and the wheelchair motion directions. Thus, the

proposed method represents a novel technique to

acquire the variety data that has been difficult to get

using other systems.

Figure 1: Outline of the proposed method.

3 PROPOSED METHOD

Figure 2 shows the steps followed to develop the

proposed method. In the following, we first describe

the development of the detection model, including

dataset creation and the design of the model). We then

describe the development of the tracking model,

including the tracking model design, camera

calibration, and the calculations used to acquire the

position and direction information.

Figure 2: Steps to develop the proposed method.

3.1 Detection Model

Here, we describe the development of the model used

to detect a wheelchair player in the acquired images.

Previous studies have reported monocular camera-

based tracking methods that use the YOLO method

(Redmon, 2016) to detect a bounding box (i.e., a

square region around each player in the image) (Buric,

2019) (Zhang, 2020). However, these previous

studies were limited to able-bodied athletes. When

Deep learning

Output

Tracking data

(Position, Direction)

Input

Video

1. Detection model

Dataset creation

Detection model design

2. Tracking model

Tracking model design

Camara calibration

Position and direction calculation

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

180

applied to a wheelchair player, bounding boxes were

only detected for the player, i.e., the wheelchair was

not included. Thus, the positions of the wheelchair

players were identified as being above the ground,

and correct positions could not be detected. Therefore,

an effective model to accurately detect wheelchairs is

required.

In previous studies (Buric, 2019) (Zhang, 2020),

the center of the bounding box or the midpoint of the

lower edge of the bounding box was used as the

player position. In the current study, we developed a

method to estimate the wheelchair structure by

applying a human posture estimation model. Based

on this model, the bottom points of each wheel are

detected, and the midpoint of the bottom points of

each wheel is calculated as the wheelchair player’s

position (shown in Figure 3). This method can detect

the wheelchair player’s positions using a monocular

camera independent of the camera positions,

wheelchair directions, and wheelchair player’s

posture.

Figure 3: Wheelchair and player detection by the proposed

model.

3.1.1 Dataset Creation

To the best of our knowledge, methods to estimate

wheelchair structure from a camera image have not

been reported. Thus, we developed a model to

estimate the wheelchair structure. A marker-less

posture estimation technique that uses a camera

image requires a large dataset to optimize a large

number of parameters. Thus, developing a wheelchair

structure model from scratch required a large dataset

of images with corresponding wheelchair key point

coordinates. However, it is difficult to collect a large

number of images of a specific category, e.g.,

wheelchair sports. In addition, it is difficult to

construct a large dataset because this requires a lot of

time and effort. Thus, in this study, we adapted a

retraining method (Dai, 2015) that converts the

human pose estimation model trained on the MS

COCO library (Lin, 2014), which is a large human

pose dataset. A new wheelchair structure estimation

model can be created even with a small wheelchair

sports dataset. In this study, we constructed a

wheelchair sports dataset containing the feature

points of wheelchair players and their wheelchairs to

be used in the retraining method.

Figure 4 shows the key point coordinates. Here, the

key points included the facial parts and the upper

body joint points, in reference to the MS COCO data

used in the pretraining process. The wheelchair key

points were the centers of the left and right wheels

and the bottoms of each wheel, which are common to

all wheelchairs and can be used to capture the

structure of the wheelchair effectively. In this study,

a total of 17 key points (i.e., nose, left eye, right eye,

left ear, right ear, left shoulder, right shoulder, left

elbow, right elbow, left wrist, right wrist, left hip,

right hip, center of left wheel, center of right wheel,

bottom of left wheel, and bottom of right wheel) were

defined in the wheelchair sports dataset. The knees

and ankles defined in the MS COCO dataset were

replaced by the centers of each wheel and the bottoms

of each wheel in the wheelchair sports dataset. These

changes were implemented to facilitate efficient fine

tuning of the parameters. The process used to

construct the wheelchair sports dataset is described as

follows.

Step 1. Automatically collect (royalty-free)

wheelchair sports images from the Internet.

Step 2. Normalize the image resolution (to 640 × 380

dpi).

Step 3. Mask people in the images who were not

related to the wheelchair players (e.g.,

referees and spectators).

Step 4. Annotate the 17 key points.

Figure 4: Key point coordinates．

In total, the wheelchair sports dataset contained

approximately 2300 images with approximately 6000

subjects. The dataset included images of wheelchair

basketball, rugby, tennis, badminton, and track and

field. The images were collected from various

wheelchair players in terms of gender and ethnicity.

The pixel size and posture of the wheelchair players

Pro

osed metho

Conventional method

left shoulder

nose

left eye

right eye

left ear

right ear

left elbow

right elbow

left wrist

left hip

right shoulder

right wrist

center of left wheel

right hip

bottom of left wheel

center of right wheel

bottom of right wheel

Development of Monocular Vision-Based Tracking Method for Wheelchair Sports

181

in the images also differed, and some of the images

included overlapping wheelchair player images. The

key point coordinates on the images were annotated

manually by sports biomechanics experts. In addition,

the data were divided randomly into training and test

sets at a ratio of 7:3.

3.1.2 Detection Model Design

We adapted a human posture estimation model

pretrained on the large-scale MS COCO dataset (Lin,

2014), and we retrained it on the acquired wheelchair

sports dataset. As the foundation model, we used the

Mask R-CNN (He, 2017), which is a widely used,

flexible, and generic framework for human posture

estimation methods. The architecture of the

foundation model is shown in Figure 5. As shown,

this model comprises three networks, i.e., the

backbone network to extract features from the RGB

images, the region proposal network to detect the

regions of players and wheelchairs, and the key point

branch to extract the key point coordinates of the

players and wheelchairs. Thus, fine tuning the

parameters of the three networks was required in this

study.

First, the initial parameter weights were obtained

by pretraining the algorithms on the MS COCO

dataset, which contains posture information for

approximately 150,000 humans in approximately

60,000 images. Then, the training data from the

acquired wheelchair sports dataset were used for fine

tuning. The optimization function was the Adam

optimizer (Kingma, 2014) with a learning rate of 0.01.

Here, 30% one of the training data was used as

validation data, and the parameter weights with the

minimized loss in the validation data were selected.

The source code was implemented in Open CV,

Python, and PyTorch, and training was performed

using an NVIDIA Tesla V100 GPU on Google

Colaboratory.

As a result, a new posture estimation model was

developed that outputs the key point coordinates of

the wheelchair player. Note that only the coordinates

Figure 5: Architecture of the fundamental model.

of the bottoms of each wheel were used in the tracking

method.

3.2 Tracking Model

3.2.1 Tracking Model Design

In Section 3.1, we described the model used to detect

the key point coordinates (i.e., the coordinates of the

bottoms of each wheel) in the images. However, this

model was insufficient for the overall task due to

occlusion caused by players overlapping images or

motion blur caused by quick movements. Thus, we

employed a model that tracks the bottoms of each

wheel of the same player’s and corrects the missing

frames by linking a series of detection results between

video frames. Here, we used the Byte Track method

(Zhang, 2022) to track multiple objects. Byte Track

links the detection results between frames by

predicting the frame-to-frame changes in key point

regions using a Kalman filter. This simple algorithm

provides high stability, high speed, and high accuracy.

The algorithm can stably track the bottom of each

wheel of the same athlete throughout the entire video.

3.2.2 Camera Calibration

In this section, we describe the process of converting

the key point coordinates (i.e., the coordinates of the

bottom of each wheel) in the detected and tracked

video image into a global coordinate (i.e., the position

in the field).

We found that there was not possible to measure

the coordinates of the calibration points in real time,

and it was impossible to enter the target space for

tracking. Thus, it was necessary to calculate the

camera parameters from the feature points in the

game or practice fields captured by the camera.

The game and practice fields have feature points

whose length and size were specified by the

International Sports Federation’s regulation. Here,

the camera parameters were calculated based on these

feature points. Figure 6 shows an image illustrating

the calculation of camera parameters on a wheelchair

tennis court. The calibrator (i.e., the court model) was

created using the court information specified by the

regulations. Using this court model, the

corresponding points of the global coordinates were

mapped to pixel coordinates (at least four points) in

the image.

The external camera parameters indicating the

camera position and orientation in three-dimensional

space were calculated using the Levenberg–

Marquardt algorithm (Moré, 1978). The internal

Backbone

network

Feature map

Region proposal

network

Fixed size

feature map

Keypoint

branch

RGB

image

Keypoint

index

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

182

camera parameters indicating the focal distance of the

camera were calculated using the hill climbing

method (Goldfeld, 1966). Using these camera

parameters, the image-based coordinate points were

converted to global coordinate points, and by

exchanging the court model, it is possible to adapt the

algorithm to other sports.

The camera parameters were calculated using a

single frame in the video; thus, the method did not

support cases where the external camera parameters

changed in the same video (e.g., camera pan, tilt,

zoom, and position shift).

Figure 6: Camera calibration.

3.2.3 Positions and Directions Calculation

The two-dimensional (2D) position of each

wheelchair player on the field and the corresponding

wheelchair directions were calculated from the global

coordinates of the key points (i.e., the bottom of each

wheel). Figure 7 shows the coordinate frames in the

case of wheelchair tennis. The 2D position of the

wheelchair player



𝑥



, 𝑦





was calculated as the

midpoint of the bottom of each wheel.



𝑥



, 𝑦









𝑥



𝑥



 2

⁄

, 𝑦



𝑦



 2

⁄ 

(1)

The wheelchair direction angle θ was calculated as

follows. This represented the sagittal plane angle of

the wheelchair.

𝜃  𝑎𝑟𝑐𝑡𝑎𝑛𝑥



𝑥



/𝑦



𝑦





(2)

Figure 7: Example of global coordinate and direction in

wheelchair tennis.

4 RESULTS

4.1 Accuracy of Pose Estimation Model

We evaluated the accuracy of the detection models

developed (Section 3.1) on the wheelchair sports

dataset using the test data, which were not used for

training. Table 1 shows the error of each model. Here,

the unit is pixels. The results for the person’s posture

are the average of the errors for each key point (i.e.,

eyes, nose, ears, shoulders, elbows, wrists, and hips),

the results for the wheelchair structure are the average

of the errors for each key point (i.e., the centers and

bottoms of each wheel), and the results for the person

and wheelchair structure are the average of the errors

for all key points. For the human key point, the mean

absolute error (MAE) was 4.43 pixels. The widely

used methods for human posture estimation, Mask R-

CNN (He, 2017) and Open Pose (Cao, 2021), were

4.68 and 4.51 pixels. Thus, the MAE value obtained

by the proposed method was greater than that of the

existing methods. The proposed method improved the

estimation error by more than 1.7% compared to the

existing methods (He, 2017) (Cao, 2021). For the

wheelchair key points, the MAE was 6.22 pixels.

These results confirm that the proposed method can

be applied to various types of wheelchair sports,

scenes, and individuals, as shown in Figure 8.

Table 1: MAE between the estimated coordinates and

manually annotated coordinates (unit: pixels).

Human

ose

Wheelchair

ose

Human an

heelchai

Mask R-CNN

(He, 2017)

4.68 - -

OpenPose (Cao, 2021) 4.51 - -

Proposed metho

4.43 6.22 4.94

Figure 8: Examples of estimation results by the proposed

method (cropped to focus on wheelchair and humans).

↓

RGB ima

e Court model →

↑ Calculated

camera

parameters

Corresponding point

to global coordinates

(

-5.49

11.89

)

(

5.49, 11.89, 0

)

(

-5.49, 0, 0

)

(

5.49, 0, 0

)

𝑥

𝑦

360 𝑑𝑒𝑔

5.49, 0

0, 11.89

0, 0

0 𝑑𝑒𝑔

𝑥



, 𝑦





𝑥



, 𝑦





𝑥



, 𝑦





𝜃

Development of Monocular Vision-Based Tracking Method for Wheelchair Sports

183

4.2 Accuracy of Tracking Model

4.2.1 Data Collection

To evaluate the accuracy of the tracking model, an

experiment was conducted during the wheelchair

tennis matches. Six elite Japanese tennis players

participated in the study. The matches were held on

an indoor tennis court, and the players used the same

wheelchairs they typically use in competitions. The

players were divided into three groups and played one

game (singles match). The players were requested to

play with the same intensity as in international

competitions. Two cameras (Pocket Cinema Camera

4K by Blackmagic Design Pty. Ltd., Port Melbourne,

Australia) were placed at each corner of the tennis

court. The height of the camera position was

approximately 6 m. Each camera monitored half of

the court. The resolution was 4K, and the frame rate

was 60 fps. Present study was conducted in

accordance with the Declaration of Helsinki, and the

protocol was approved by the Ethics Committee of

Nagaoka University of Technology.

4.2.2 Accuracy of Player Detection

The tracking data of the wheelchair players were

output from the video images using the proposed

method. Figure 9 shows an example of the tracking

result. The player’s trajectory was overlaid on the

input image. Table 2 shows the detection rate of each

wheelchair player. As can be seen, the proposed

method was able to detect the wheelchair players in

all frames.

Figure 9: Image of the tracking results.

Table 2: Detection success rate of wheelchair player using

the proposed method.

Total data [s]

Detection

data

[

]

Detection rate

[

]

Pla

er 1 2650 2650 100

Pla

er 2 2650 2650 100

Pla

er 3 2300 2300 100

Pla

er 4 2300 2300 100

Pla

er 5 2900 2900 100

Pla

er 6 2900 2900 100

All 15700 15700 100

4.2.3 Accuracy of Player Position

The positions of the wheelchair players were

calculated using the proposed method. The videos

were also digitized manually as reference values for

validation. Here, for each player, 120 frames were

selected randomly, and the bottoms of each wheel

were digitized manually. The midpoint of each wheel

was taken as the true value, and the coordinate

transformation by the camera calibration was the

same the proposed method. Table 3 shows the

position determination errors of the proposed method

(coordinate frames - according to Figure 7). The

MAE in the horizontal direction (x) was 0.03 m, in

the depth direction (y) was 0.10 m, and the root mean

square error (RMSE) was 0.11 m. The values of "All"

in Table 3 were calculated from all data of all players.

Table 3: Position determination errors of our method.

MAE x

[

]

MAE

[

]

RMSE

[

]

Pla

er 1 0.03 0.09 0.10

Pla

er 2 0.03 0.09 0.10

Pla

er 3 0.03 0.14 0.15

Pla

er 4 0.02 0.09 0.09

Pla

er 5 0.03 0.09 0.10

Pla

er 6 0.03 0.12 0.12

All 0.03 0.10 0.11

4.2.4 Accuracy of Wheelchair Direction

The wheelchair motion directions of the players were

calculated using the proposed method. As in the

evaluation of the positional errors, here, the true

values of the wheelchair directions were calculated by

digitized manual data. Table 4 shows the wheelchair

directions errors of the proposed model, and the

coordinate system is shown in Figure 7. As can be

seen, the mean ± SD was −2.23 ± 8.57 degrees, and

the MAE was 6.78 degrees. The values of "All" in

Table 4 were calculated from all data of all players.

Table 4: Wheelchair direction error of our method.

Mean ± SD

[

MAE

[

Pla

er 1 −3.22 ± 8.65 7.82

Pla

er 2 −2.48 ± 9.89 8.02

Pla

er 3 −2.80 ± 8.71 7.13

Pla

er 4 −0.40 ± 7.98 6.34

Pla

er 5 −0.73 ± 8.32 6.61

Pla

er 6 −3.75 ± 7.05 6.46

All −2.23 ± 8.57 6.78

trajectory over the

ast 0.5 seconds

RGB ima

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

184

5 DISCUSSIONS

We found that the proposed method estimated the

wheelchair structure with the same accuracy as

existing human posture estimation models, as shown

in Table 1. Note that the constructed wheelchair

sports dataset includes images of wheelchair

basketball, rugby, tennis, badminton, and track and

field; thus, the proposed method provides tracking

data for a wide variety of wheelchair sports. In

addition to wheelchair structure, the proposed model

estimates the player’s upper body posture. In addition

to measuring the wheelchair player positions, it is also

possible to analyze the upper body movements. For

example, the proposed method could be used to

evaluate the wheelchair rowing motion. Also, it

would be possible to analyze the relationship between

the upper body usage and chair work skill using the

proposed method.

As shown in Table 2, the detection success

accuracy for wheelchair players were 100%. In a

previous study (Sarro, 2008), the detection

successrate of wheelchair rugby players using a video

camera was approximately 74%, and that of

wheelchair soccer players was approximately 94%.

The proposed method demonstrates higher accuracy

than previous study, and it achieved the accuracy

required for use in sports.

The RMSE of the positions determinated by the

proposed method was 0.11 m, as shown in Table 3.

When using an LPS in wheelchair sports, the MAE

was 0.19–0.32 m in wheelchair rugby and wheelchair

basketball, respectively (Rhodes, 2014), and the

MAE was 0.37 m for wheelchair tennis (Perrat, 2015).

The results obtained for the proposed method indicate

that it outperforms these existing methods in terms of

accuracy.

In this study, the position accuracy was evaluated

in wheelchair tennis. The proposed method can be

applied to other wheelchair sports by exchanging the

court model for camera calibration. The proposed

method is an innovative tracking system that does not

require base stations or devices attached to the players,

and it can realize high position detection accuracy

using only a single camera.

The MAE of the wheelchair directions tracked by

the proposed method was 6.78 degrees, as shown in

Table 4. Methods based on a single IMU sensor are

widely used to measure wheelchair directions. For

example, previous studies reported 8.1 degrees (van

Dijk, 2022) and 11.0 degrees (Rupf, 2021). Thus, the

proposed method outperforms methods based on a

single IMU sensor. In addition, to the best of our

knowledge, the proposed method is the first based on

a monocular camera. Thus, the proposed method

provides a simplified novel tool to obtain kinematic

data for wheelchair sports.

The proposed method provides the movement

information of wheelchair players using a single

camera placed at the side of the field or near audience

seats. Thus, it is useful for training load management

and evaluating on-court performance. In addition, for

competitive sports, the proposed method can be used

to acquire kinematic data of opponents to improve the

analysis of tactics.

The proposed method can be applied to the analysis

of past legendary players and to compare the past and

current performance of the same player, even if it is

not possible to acquire new data using the tracking

system. We believe that our findings contribute to the

quantitative performance evaluation of wheelchair

athletes.

Finally, we describe the limitations observed in this

study. We found that the proposed method exhibits a

larger error in the depth direction than in the

horizontal direction due to the single camera (Table

3). In addition, the position error increases when the

number of pixels per wheelchair player decreases.

Thus, with the proposed method, it is necessary to

consider the image acquisition conditions. In this

study, the evaluation was conducted for singles

wheelchair tennis; thus, the accuracy may decrease

according to the overlap of wheelchair players.

Therefore, in the future, we plan to evaluate the

proposed method when multiple players are present

on the same field.

6 CONCLUSIONS

In this study, we developed a tracking method to

measure the kinematic data of wheelchair sports using

a monocular camera. With the proposed method, the

RMSE of the wheelchair player position was 0.11 m,

and the MAE of the wheelchair direction was 6.78

degrees. In addition, the proposed method achieved

higher accuracy than existing tracking methods, e.g.,

the LPS and IMU sensor–based methods. The

proposed method provides a simple tool to obtain

kinematic data in wheelchair sports, which have not

been collected previously. This research contributes

to the quantitative performance evaluation of

wheelchair athletes.

Development of Monocular Vision-Based Tracking Method for Wheelchair Sports

185

ACKNOWLEDGEMENTS

This work was supported by the "Functional

Development Project for Resilient Athlete Support"

of Japan Sports Agency.

REFERENCES

Buric, M., Ivasic-Kos, M., & Pobar, M. (2019). Player

tracking in sports videos. Proceedings of the

International Conference on Cloud Computing

Technology and Science, 334–340.

Cao, Z., Hidalgo, G., Simon, T., Wei, S. E., & Sheikh, Y.

(2021). OpenPose: Realtime Multi-Person 2D Pose

Estimation Using Part Affinity Fields. IEEE

Transactions on Pattern Analysis and Machine

Intelligence, 43(01), 172–186.

Dai, A. M., & Le, Q. V. (2015). Semi-supervised Sequence

Learning. Advances in Neural Information Processing

Systems, 28.

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

Maximization by Quadratic Hill-Climbing.

Econometrica, 34(3), 541.

Grogan, A. (2012). Paralympic technology. Engineering

and Technology, 7(8), 28–31.

Halson, S. L. (2014). Monitoring Training Load to

Understand Fatigue in Athletes. Sports Medicine, 44(2),

139–147.

He, K., Gkioxari, G., Dollar, P., & Girshick, R. (2017).

Mask R-CNN. In Proceedings of the IEEE

International Conf. on Computer Vision, 2961–2969.

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

Stochastic Optimization. 3rd International Conference

on Learning Representations.

Laferrier, J. Z., Rice, I., Pearlman, J., Sporner, M. L.,

Cooper, R., Liu, T., & Cooper, R. A. (2012).

Technology to improve sports performance in

wheelchair sports. Sports Technology, 5(1–2), 4–19.

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

Ramanan, D., Dollár, P., & Zitnick, C. L. (2014).

Microsoft COCO: Common objects in context.

European Conf. on Computer Vision, 8693, 740–755.

Linke, D., Link, D., & Lames, M. (2020). Football-specific

validity of TRACAB’s optical video tracking systems.

PLOS ONE, 15(3), e0230179.

Mason, B. S., Porcellato, L., Van Der Woude, L. H. V., &

Goosey-Tolfrey, V. L. (2010). A qualitative

examination of wheelchair configuration for optimal

mobility performance in wheelchair sports: a pilot study.

Journal of Rehabilitation Medicine, 42(2), 141–149.

Moré, J. J. (1978). The Levenberg-Marquardt algorithm:

Implementation and theory. Numerical Analysis, 105–

116.

Pansiot, J., Zhang, Z., Lo, B., & Yang, G. Z. (2011).

WISDOM: wheelchair inertial sensors for displacement

and orientation monitoring. Measurement Science and

Technology, 22(10), 105801.

Perrat, B., Smith, M. J., Mason, B. S., Rhodes, J. M., &

Goosey-Tolfrey, V. L. (2015). Quality assessment of an

Ultra-Wide Band positioning system for indoor

wheelchair court sports. Proceedings of the Institution

of Mechanical Engineers, Part P: Journal of Sports

Engineering and Technology, 229(2), 81–91.

Perret, C. (2017). Elite-adapted wheelchair sports

performance: a systematic review. Disability and

Rehabilitation, 39(2), 164–172.

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016).

You Only Look Once: Unified, Real-Time Object

Detection. In Proceedings of the IEEE Conference on

Computer Vision and Pattern Recognition, 779–788.

Redwood-Brown, A., Cranton, W., & Sunderland, C.

(2012). Validation of a real-time video analysis system

for soccer. International Journal of Sports Medicine,

33(8), 635–640.

Rhodes, J., Mason, B., Perrat, B., Smith, M., & Goosey-

Tolfrey, V. (2014). The validity and reliability of a

novel indoor player tracking system for use within

wheelchair court sports. Journal of Sports Sciences,

32(17), 1639–1647.

Rupf, R., Tsai, M. C., Thomas, S. G., & Klimstra, M. (2021).

Original article: Validity of measuring wheelchair

kinematics using one inertial measurement unit during

commonly used testing protocols in elite wheelchair

court sports. Journal of Biomechanics, 127.

Sarro, K. J., Misuta, M. S., Burkett, B., Malone, L. A., &

Barros, R. M. L. (2010). Tracking of wheelchair rugby

players in the 2008 Demolition Derby final. Journal of

Sports Sciences, 28(2), 193–200.

Shepherd, J. B., James, D. A., Espinosa, H. G., Thiel, D. V.,

& Rowlands, D. D. (2018). A Literature Review

Informing an Operational Guideline for Inertial Sensor

Propulsion Measurement in Wheelchair Court Sports.

Sports 2018, 6(2), 34.

an der Slikke, R. M. A., Berger, M. A. M., Bregman, D. J.

J., & Veeger, H. E. J. (2016). From big data to rich data:

The key features of athlete wheelchair mobility

performance. Journal of Biomechanics, 49(14), 3340–

3346.

van der Slikke, R. M. A., Mason, B. S., Berger, M. A. M.,

& Goosey-Tolfrey, V. L. (2017). Speed profiles in

wheelchair court sports; comparison of two methods for

measuring wheelchair mobility performance. Journal

of Biomechanics, 65, 221–225.

van Dijk, M. P., van der Slikke, R. M. A., Rupf, R.,

Hoozemans, M. J. M., Berger, M. A. M., & Veeger, D.

J. H. E. J. (2022). Obtaining wheelchair kinematics with

one sensor only? The trade-off between number of

inertial sensors and accuracy for measuring wheelchair

mobility performance in sports. Journal of

Biomechanics, 130, 110879.

Zhang, Y., Chen, Z., & Wei, B. (2020). A Sport Athlete

Object Tracking Based on Deep Sort and Yolo V4 in

Case of Camera Movement. 2020 IEEE 6th

International Conf. on Computer and Communications,

1312–1316.

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

Luo, P., Liu, W., & Wang, X. (2022). ByteTrack: Multi-

object Tracking by Associating Every Detection Box.

European Conf. on Computer Vision, 13682, 1–21.

icSPORTS 2023 - 11th International Conference on Sport Sciences Research and Technology Support

186