You are on page 1of 11

【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

허프변환과 YOLO 기반의 골프공 궤적 추적


Tracking The Trajectory of a Golf Ball Based on Hough Transform
and YOLO

1)
런자이 바이다, 이상웅
Ranjai Baidya, Sang-Woong Lee

(13120) 경기도 성남시 수정구 성남대로 1342 가천대학교 IT융합공학과, AI ‧ 소프트웨어학부


[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

ranjai123baidya@gmail.com, slee@gachon.ac.kr

요 약

드라이브나 티샷을 할 때 육안을 통해 수동적으로 골프공의 궤적을 시각화하는 것은 어려운 작업이 될 수 있는데,
그 이유는 공의 크기가 작고 공이 이동하는 속도가 빠르기 때문이다. 이러한 작업의 특성으로 인해 값비싼 센서 및
구성 요소를 사용하여 제조된 복잡한 장치가 항상 요구되어 왔습니다. 본 논문에서는 골프 티샷중 골프공의 궤적을
추적하는 시스템을 제안한다. 제안하는 시스템은 기존의 컴퓨터 비전 기술과 현대적인 심층 신경망을 결합하여 모노
스테레오 비디오에 궤적을 추적한다. 골프 볼 추적에 앞서, 골프공과 플레이어의 위치를 특정 짓기 위해
YOLO(You Only Look Once) V3 딥 신경망 모델을 사용한다. 우리는 클럽과 공 사이의 초기 충격 지점을 추적
하기 위해 허프 변환을 사용한다. 또한 볼을 추적하기 위해 차영상을 이용하여 비디오의 모션 정보를 추출한다. 제
안하는 시스템의 효과를 검증하기 위해 다수의 비디오에서 테스트되었으며 각각의 결과를 보여준다.
www.earticle.net
Abstract

Manually visualizing the golf ball's trajectories during a drive or a tee shot could be an unsettling
task for viewers, the reason being the small size of the ball and the high speed at which it travels.
This nature of the job has always required intricate systems manufactured using an expensive set of
sensors and components. In this paper, we propose a system to draw the trajectory of a golf ball
during a drive. The proposed system combines the classical computer vision techniques and a modern
deep neural network to project the trajectories over videos taken from monocular cameras. Before
starting to track the ball, the system uses YOLO(You Only Look Once) V3, deep neural network
model to perform localization of the golf ball and the player. To track the initial point of impact
between the club and the ball the system uses Hough transform. Finally to track the ball, the system
extracts motion information of the video by using frame differencing. Unto the information received
from frame differencing, the system applies filters based on a few assumptions, size and direction of
movement to obtain the detections of the golf ball. The proposed system was tested on a multitude of
videos to verify its effectiveness and the respective results are presented.

1) corresponding author

42

논문투고일자 : 2021. 03. 28 심사일자 : 2021. 03. 29 게재확정일자 : 2021. 04. 23


【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

키워드: 골프 공, 추적, 차영상, 임계값, 허프 변환, 욜로-V3


Keyword: Golf ball, Tracking, Frame Differencing, Threshold, Hough Transform, YOLO-V3

1. Introduction case. Firstly, the golf balls' size is so small


that the field of view of a camera will not even
Golf is a sport adored by an abundance of
be able to capture the ball in half of the entire
people. Along with having a vast magnitude in
time the ball is in flight. Secondly, the high
terms of the number of players worldwide, the
velocity of the ball may make it look blurry in
sports viewers' numerical figure is more
the video captured. Furthermore, there are
comprehensive. However, even if it is popular
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

high chances of occlusion of the ball. Considering


among huge masses, the viewing experience,
these factors using just the modern deep
the overall process of analyzing the player's
learning-based tracking algorithm is out of
performance, and visualizing the trajectories
the question. If we were to use just those
can be tedious and daunting. To nullify this,
state-of-the-art deep learning-based tracking
tools like launch monitors have been fabricated
algorithms, likely, whatever tiny details the
and made available for the public. They assist
video could offer regarding the moving ball
in keeping track of a golf ball's motion
would be lost, and the model would perform
parameters like velocity, spin, launch angle,
poorly.
travel distance, and maximum elevation during
Mathematical modeling of the golf ball
its drive. Some of these products also have
trajectory is another alternative that could be
features that can help visualize the golf ball's
used in our application. An accurate model can
drive's clear trajectory. All of this has been
deliver the desired result. But the aerodyna-
www.earticle.net
possible with the use of complex and expensive
mics of golf balls is a highly complex subject
components within these devices. The commercially
matter. Unlike most of the other objects
available tools are primarily based on multiple
globally, a golf ball's trajectory does not
sensors like dual cameras, a radar camera, and a
simply follow a standard parabolic path. The
camera laser. While these costly multi-sensor
golf ball's dimples bring in factors like lift and
types of equipment tend to be highly accurate,
drag to its aerodynamics, and the trajectories
they are also mostly out of the general public's
are complicated to predict. A precise mathe-
reach as they come at an extreme price that
matical model of golf ball aerodynamics that
not everyone can afford. We aim to develop a
factors in all the possible parameters that
simpler inexpensive system that can be a
affect its trajectory is yet to be designed. It
cheaper alternative to picturing the golf ball
may not come to fruition shortly[1, 2].
drives and tee shots trajectories.
Nonetheless, even if we were to use a
In an ideal world scenario, just a high-
mathematical model that is as close as possible
resolution camera would have been sufficient
to reality, measuring all the parameters like
to capture the entirety of a golf ball drive
ball velocity, spin, and launch angle, which
trajectory; however, in reality, that is not the

43
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

shall be used in the model, is a difficult task. Lots of effort have also been put into
This task would require equally complex mathematically modeling the aerodynamic
systems as previously discussed and will defy properties of a golf ball. While systems are
the purpose of our work. using videos from monocular camera to track
In the current digital era, the most easily other sports, not many have attempted to do
accessible resources for any topic are the the same with golf.
images and videos available on the internet. It
2.1 Mathematical Modeling and Experimen-
is no exception in terms of golf as well. Golf
tation with Physical Parameters
drive videos are abundant on the web that can
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

be useful for analyzing golf shots' trajectories In the past, there have been attempts to
are videos of the golf shots. To make the define a golf ball's aerodynamics using mathe-
visualization of the course simple, we use the matical models and experimentation. Previously,
videos of golf shots taken from monocular there was an investigation regarding simulation
camera. The decision to use 2D videos of a golf ball's trajectory using the initial
captured using a single camera system speed, rotation of the ball, and wind speed,
restrains us from much information. Unlike assuming linear impact of the velocity on the
methods using multiple cameras or even the drag force and independence of motion along
systems with additional sensors like radar or each axis with each other [4]. Another study
laser, our system will not have access to the proposed an even better model by including
information like velocity, spin, and launch the law for spin rate decay and non-linear
angle of the ball. Hence, to make the most of dependence of drag on Reynolds number [5].
the tiny amount of information available to our There was some niche study related to the
system, we have implemented a www.earticle.net
combination parameters of motion of golf. A wind tunnel
of old fashioned computer vision techniques technique was developed to quantify the
like frame-wise differencing and currently aerodynamic forces affecting golf balls while
trending deep neural network-based tracking varying the Reynolds number and the spin
algorithm YOLO V3 [3]. We also make few rate. The measured aerodynamic data was
scenario-specific assumptions that will help then used to calculate the trajectories and
us in obtaining a more robust and efficient compare them to real-world data [6]. After
system. measuring the drag coefficients of variety of
commercially available balls under varying
2. Related Works wind speeds, a study found out experimentally
that the drag coefficients varied significantly
There have been quite a few attempts to
due to the variation of number and sizes of
track a golf ball's flight and plot it into a graph
dimples on the ball [1]. The flow pattern of a
for visualization. Systems are using classical
golf ball with rotation was studied computationally
object tracking methods, modern computer vision
in a Large-eddy simulation [7]. The effect of
approach, and also sensor-based techniques.
wind on aerodynamics has been reviewed and

44
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

presented accordingly [8]. A new golf flight players [14]. An extended short-term memory
model was also suggested to be used in a golf network(LSTM) based recurrent neural network
game simulation in another approach [9]. system was used to perform extrapolation of
a golf ball's remaining trajectory. The golf ball's
2.2 Sensors-Based Approaches
three-dimensional position is known for most
Implementation of systems using sensors on of the duration the ball is in flight [15]. A
the golf equipment or measuring the club's database containing many 3D trajectories of
movement parameters and the ball are also golf balls captured using a dual-camera system
interesting. A sensor-integrated golf club was was used to match and construct a 3D course
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

suggested. Two orthogonal affixed train gage of the golf ball. Many models were tried to
sensors, a 3-axis gyroscope, and an accelero- help in the scenarios when one of the cameras
meter were used to monitor the club's in a dual-camera system failed [16]. The golf
movement during a golf swing [9]. club's movement-related information was
A traditional computer vision embracing used to distinguish the golf shot trajectories
system was designed using two calibrated stereo into one of nine predefined shot patterns using
cameras, where a frame differencing approach neural network and support vector machine
was implemented in a video with an unchanged models [17].
background [10]. Studies have also been done
2.4 Other Sports-Related Works
for creating virtual golf games using multiple
sensors. A technique was suggested to auto- A system using a camera was devised to
matically detect high-speed flying ball from track a tennis ball and predict the part of the
multi-exposure images under varying light trajectories where the ball could not be tracked.
conditions [11]. A new type of www.earticle.net
marker was For prediction, the system used a simple
defined to be used motion capture system physical model considering a parabolic arc and
[12], a concept similar to this can be imple- factoring just the gravity [18]. A similar approach
mented using this, however our system shall was also suggested to track a volleyball from
use just the features that can be obtained from a single camera volleyball sequence and
videos captured from monocular cameras. create a 3D trajectory approximation [19].
Another system was devised to study basketballs'
2.3 Approaches Based on Machine Learning
trajectories using monocular video sequence
In a recent approach, an online golf ball by assimilating some basketball-related domain
detection and tracking system was suggested, knowledge and real-world physical attributes
which used the combination of Kalman filter of the ball motion [20]. A different ball
and convolutional neural networks like YOLO tracking algorithm was used to track a baseball
v3 and its tiny version, YOLO v4, Faster using a Kalman filter in a 2D video sequence
R-CNN, SSD, and RefineDet. [13]. A study [21].
was used track people in surveillance images,
a similar idea can be used to detect golf

45
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

3. Methods video generally means there is a difference


between corresponding pixel values in the
The proposed system to track the trajectory
consecutive frames. This concept is the basis
of a moving golf ball comprises applying
of the frame differencing approach adapted to
frame-wise differencing and thresholding
distinguish the moving parts in captured videos.
along with a deep learning based detection
The values of each corresponding pixel of the
algorithm YOLO V3. The system utilizes video
consecutive frames are subtracted from one
sequences of golf shots captured from any
another, and the result is compared to a
monocular camera setup.
threshold value to get a binary image. Ideally,
The system, initially, applies frame differencing
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

this would return an image that highlights the


and thresholding to frame by frame obtain the
pixels that are moving in the video. Never-
coordinates of the pixels that are changing in
theless, noise is prevalent in this approach and
the video. The coordinates of these changing
hence extracting the objects of interest may
pixels in turn can be processed by the system
not be as straight forward as in theory.
to obtain necessary information, using a series
of hypothesis and also techniques like Hough 3.2 YOLO-V3
transform[22] and YOLO V3. The information
YOLO discarded the previously favored
obtained by the system are position of the
approach of using classifiers for performing
moving golf club, the position of initially static
detections. It represented frame object detection
golf ball, the ideal beginning frame to start
as a regression problem, where spatially distant
tracking the ball and then later on the overall
bounding boxes are related to the probability
positions of the golf ball throughout its
of belonging to a particular class. In YOLO, a
trajectory. www.earticle.net
single neural network predicts both the bounding
In this section, the techniques, hypothesis
boxes and the class probabilities directly from
and the step wise explanation of processes
the whole of the images in a single go. The use
used in this system have been explained. At
of a single network for the whole pipeline
first the techniques like frame differencing and
enabled the optimization of detection performance
thresholding, YOLO V3 and Hough transform
[22]. YOLO-V3 is an improvement on the
are described in the sections 3.1, 3.2 and 3.3
original YOLO. The YOLO-V3 model features
respectively. The detailed process of obtaining
multi-scale detection, a more robust feature
the required informations and tracking the
extractor network, and a slightly different loss
trajectory of the golf have been described in
function. As compared to the previous model,
sections 3.4, 3.5 and 3.6.
the YOLO-V3 model can detect more variation
3.1 Frame-wise Differencing and thresholding of targets from big to small [3].

The movement of objects captured in a video 3.3 Lines Detection using Hough transform
or the camera's movement while capturing a
Hough transform is an alternative repre-
video or the presence of noise in the captured
sentation of the coordinate system that can d

46
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

Fig. 1 General System Overview

etect lines in an edge image. In the original mapped in this new space, a lot of cosine
Hough space, the two axes are the slope and curves will be generated. The curves from the
the x-intercept of a line represented by the closer edge points shall intersect one another
equation of the form, y = mx + c. A line from on a specific value of (ρ, θ) pair. Finding the
www.earticle.net
an edge image can be described as a point in (ρ, θ) pairs having several intersections
the Hough space, while a single edge point in greater than a threshold line can be detected
the standard coordinate system is a whole set using Hough's transform [22, 24]. It is
of lines passing through the point. However, necessary to note that the image passed on to
the Hough space cannot detect vertical lines Hough's transform needs to be an edge image.
due to the infinite value of slope in such cases. So, in our system, we apply the canny edge
To avoid this, a straight line is represented by detection algorithm before performing Hough
a line called the normal line that passes transform over it.
through the origin and perpendicular to that
3.4 Determination of the ideal frame number
straight line. The equation representing this
to start tracking the ball
normal line is ρ= x cos(θ) + y sin(θ) where
ρ is the length of the normal line, and θ is the Naturally, a golf player takes up three
angle between the normal line and the x-axis. different positions before driving the golf ball.
Now the Hough space's axes shall be ρ and In the proposed system, an assumption has
θ, and an edge point produces a cosine curve been made that the three positions are always
in the Hough Space. If all the edge points are taken up by the player in the video. These

47
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

Fig. 2 Different positions taken up by a golf player before hitting the golf ball and result of image
differencing on consecutive frames of images in the corresponding time.

three positions can be observed in fig 2. The degrees with the x-axis(anti-clockwise direction)
first position is when the player aims the club and appearing nearer to the top half of the
to get the job right for suitable contact. The player's body. The golf player is detected
second position is assumed when the player using YOLO V3.
lifts the club to get enough momentum to drive To find the lines in the images, we apply
the golf ball. The final position is when the Hough's line algorithm described in section 3.3
club is just about to hit the ball. So, we look is used in the differenced pictures, and to find
for frames where the player takes www.earticle.net
up these the inclination of the lines, we look at the
three positions and mark the frame occurring slopes of the lines. The slope of these lines
exactly after the final position as the ideal can be found out using equation 1, where (x1,
frame to start tracking the ball. y1) and (x2, y2) are the coordinates of two
Throughout the time that the player takes end points lying on the lines found out using
up the afore mentioned positions, the golf club the Hough transform.
is constantly moving, and it appears as a line
in the differrenced frames. These lines on the   
     (1)
differenced image guide us to decide whether  
 

the conditions have occurred in any particular


frame or not. The first and the last state is The proposed system marks the frame as
indicated by a line at an inclination of 30-60º suggested by the output of this step and starts
with the x-axis(anti-clockwise direction) and tracking the ball only after the marked frame.
appearing nearer to the bottom half of the
3.5 Determination of the precise location
player's body. The second condition is denoted
of the ball when on the ground
by a line at an inclination of 120-150°

48
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

Fig. 3 Obtained Results

When the golf player aims the club before It can be debated that using only one of the
hitting the ball, the golf club held by the player two steps mentioned can be used to find the
hovers around the periphery of the ball. We ball's position. However, using only the golf
exploit this information and use the YOLO V3 club's average position would not be accurate
network to find the ball's precise location enough for our application. Also, using only
before the club hits it. Just as in section 3.4, the YOLO V3 would hinder our system when
the club position is found out using Hough there are multiple golf balls present in the
transform around the lower area of the golf video frame.
player and the slope of those lines.
3.6 Tracking the golf ball trajectory
This process is implemented onlywww.earticle.net
on the
video's initial few frames. To be more specific, After the completion of the steps mentioned
the frames used are the ones occurring in sections 3.4 and 3.5, the system would have
between the time the player starts aiming the the necessary information to start the process
ball(position 1 as mentioned in section 3.4) of tracking the trajectory of the ball. First the
and the time the player lifts the club to gain initial position of the ball is initialized to the
some momentum(position 2 as mentioned in position given by the output of the steps in
section 3.4). Throughout these frames, the section 3.4. The tracking process is initiated
club positions are averaged to estimate the from the frame marked after the processes of
coordinate where the ball could be present. section 3.5. The consecutive frames from then
This estimated coordinate acts as a guide for on are differenced and thresholding is
us to look for the pinpoint location of the ball. performed till the occurance of the last frame
Then a small 100 pixels x 100 pixels area of the video. As mentioned in the section 3.1,
around the estimated coordinate is passed to the differencing and thresholding gives us the
a pretrained YOLO V3 network to find the pin pixel positions of the moving object as well as
point location of the golf ball. the noise in the video. To find the ball among

49
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

those moving objects, we first filter the the initialized golf ball position, the stored
detections based on the area. The size of the detections are also used to find the initial
ball could be anywhere from 2 to 20 pixels equation of the lines joining those points to the
large. However, filtering only based on the detections of the consecutive frame. The new
size is not enough. There are high possibilities points satifying the threshold of 50 pixels are
of noise and other moving objects of size stacked upon previously stored detection if
similar to the ball. Hence to further filter the they belong to the same direction, otherwise
exact ball detections we use the fact that there new stacks are created if they don't belong to
is only single moving ball in each frame and any of the previous stacks. At the end the
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

while in upward or in downward motion of the length of each of these stacks are check. The
trajectory the ball needs to be moving in a stacks created by the noises and other
similar direction when observed frame after detections tend to be very short. The largest
frame. Since the videos used in this system of those stacks is then chosen as the set of
are taken from behind the players, the golf detections of the golf ball.
balls are insignificantly small when in
downward motion of the trajectory. So we try 4. Experiments and Results
only to track the ball when it is moving in
In this section, we show the details of the
upward direction.
videos and images used for the system and the
To filter the detections based on direction,
experimental results of the overall process.
the system uses detection from three
To train the YOLO v3 and test the overall
consecutive frames throughout the tracking
processes, 10707 images were extracted from
period. First the line equation are found out
different golf shot videos from YouTube. The
between the initialized ball position and www.earticle.net
detection
YOLO-V3 model was trained based on the
points of the consecutive frames. Given (x1,
transfer learning approach to distinguish just
y1) and (x2, y2) are the detections of two
the two classes, golf_player, and golf_ball.
consecutive frames, the equation of the line
After training the YOLO-V3 model, the
joining these points is given by:
network was implemented as a part of the
overall system.
          (2)
The overall system performed well, and the
results obtained on a few of the videos have
where, m is the slope of the line, can be
been presented in fig 3. The gray lines on the
computed using equation 1.
images represent the path taken by the golf
Then the distance from detections of the
ball during its flight. As can be seen from the
third consecutive frame and the lines are
output results, the system tracks the portion
calculated. If any of the distances is less than
of the trajectory where the ball is moving in
50 pixels, all three points related to these
upward direction.
distance are stored. The step is then repeated
throughout the end, however now along with

50
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

5. Conclusion namic model of a golf ball in flight.” In


Proceedings of the 1994 World Scientific
We proposed a system to track the trajectory
Congress of Golf, 1994
of the golf ball tee shots. Using techniques like
[6] P. W. Bearman, and J. K. Harvey, “Golf ball
frame differencing, thresholding, Hough transform
aerodynamics. Aeronautical Quarterly”, Vol.27(2),
along with deep neural network model YOLO-V3
pp.112-122, 1976
and some case-specific assumptions, the
[7] K. Aoki, K. Muto, and H. Okanaga, “Aerody-
proposed system performed successfully in
namic characteristics and flow pattern of a golf
tracking the golf ball to the point where it is
visible to the human eye. The system can ball with rotation.” Procedia Engineering, Vol.
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

further be improved to predict the rest of the 2, No. 2, pp.2431-2436, 2010


trajectory of the ball. [8] S. S. MALIK, and S. Saha, “Golf and Wind:
The Physics of Playing Golf in Wind.” Springer
Nature, 2021
Acknowledgement
[9] A. Umek, Y. Zhang, S. Tomažič, and A. Kos,
This work was supported by the GRRC “Suitability of strain gage sensors for inte-
program of Gyeonggi province. [GRRC-Gachon gration into smart sport equipment: A golf club
2020 (B02), AI-based Medical Information example.” Sensors, Vol. 17(4), 916, 2017.
Analysis]. [10] A. Woodward, and P. Delmas, “Computer
vision for low cost 3-D golf ball and club
tracking.” Proc. Image and Vision Computing,
References
New Zealand, pp. 11-15, 2005
[1] F. Alam, T. Steiner, H. Chowdhury, www.earticle.net
H. Moria, [11] J. S. Kim, and M. G. Kim, “Automatic high-
I. Khan, F. Aldawi, and A. Subic, “A study of golf speed flying ball detection from multi-exposure
ball aerodynamic drag.” Procedia Engineering, images under varying light conditions.”, VRCAI,
Vol. 13, pp. 226-231., 2011 2011.
[2] S. Baek, and M. Kim, “Flight trajectory of a golf [12] M. J. Kim, S. M. Lee, I. U. Haq, F. U. M.
ball for a realistic game.” International Journal Ullah, M. Y. Lee, S. W. Baik, “Study on
of Innovation, Management and Technology, Person Re-identification in Multi-view Survei-
Vol. 4, 346, 2013 llance Videos”, THE JOURNAL OF KOREAN
[3] J. Redmon, and A. Farhadi, “Yolov3: An incre- INSTITUTE OF NEXT GENERATION
mental improvement.”, arXiv preprint arXiv: COMPUTING, Vol. 17(1), pp. 16-24, 2021.
1804.02767, 2018 [13] T. Zhang, X. Zhang, Y. Yang, Z. Wang, and
[4] J.J.McPhee, and G. C. Andrews, “Effect of G. Wang, “Efficient golf ball detection and
sidespin and wind on projectile trajectory”, tracking based on convolutional neural networks
with particular application to golf. American and Kalman filter”, arXiv preprint arXiv:2012.
Journal of Physics, 56(10), pp.933-939, 1988 09393, 2020.
[5] A. J. Smits, and D. R. Smith, “A new aerody- [14] C. Lim, and J. Choi., “Unique-ID Generation

51
【논 문】 한국차세대컴퓨팅학회 논문지 2021년 4월

Marker-based Motion Capture Technology.”, [22] V. F. Leavers, “Preprocessing, Shape Detection”,


THE JOURNAL OF KOREAN INSTITUTE OF Computer Vision Using Hough Transform, pp.
NEXT GENERATION COMPUTING, Vol. 39–64
4(2), pp. 33-38, 2008 [23] J. Redmon, S. Divvala, R. Girshick, and A.
[15] A. Jansson, “Predicting trajectories of golf Farhadi, “You only look once: Unified, real-
balls using recurrent neural networks”, 2017. time object detection.”, Proceedings of the
[16] J. Sk̈old, “Estimating 3d-trajectories from IEEE conference on computer vision and
monocular video sequences”, 2015. pattern recognition, pp. 779-788.
[17] B. Bačić, “Predicting golf ball trajectories from [24] S. C. Lee, H. G. Lee, and M. K Choi, “Indoor
[Provider:earticle] Download by IP 223.39.99.207 at Thursday, January 19, 2023 4:03 PM

swing plane: An artificial neural networks path guidance system for visually impaired
approach.”, Expert Systems with Applications, persons using infrared vision.”, THE JOURNAL
Vol.65, pp.423–438, 2016. OF KOREAN INSTITUTE OF NEXT GENERATION
[18] A. Kumar, P. S. Chavan, V. K. Sharatch- COMPUTING, Vol. 6(6), pp. 24-34, 2010
andra, S. David, P. Kelly, and N. E. O'Connor,
“3d estimation and visualization of motion in
a multicamera network for sports.”, 2011 Irish 저자소개
Machine Vision and Image Processing
◆ 런자이 바이다
Conference, pp. 15-19, IEEE, September, 2011 ∙2018년 Kathmandu University Bachelors
[19] H. T. Chen, W. J. Tsai, S. Y. Lee, and J. Y. of Electrical and Electronics Engi-
Yu, “Ball tracking and 3D trajectory appro- neering
∙2021년~현재 가천대학교 나노과학기술
ximation with applications to tactics analysis 융합학과 석사과정
www.earticle.net
from single-camera volleyball sequences.”, ∙관심분야: 영상인식, 객체추적, 딥러닝
Multimedia Tools Appl Vol. 60, pp. 641–667,
2012. ◆ 이상웅
∙1996년 고려대학교 전자공학사
[20] H. T. Chen, M. C. Tien, Y. W. Chen, W. J.
∙2001년 고려대학교 컴퓨터학석사
Tsai, and S. Y. Lee, “Physics-based ball ∙2006년 고려대학교 컴퓨터학박사
tracking and 3D trajectory reconstruction ∙2007년 카네기멜론대학교 로봇공학연
구소 객원연구원
with applications to shooting location esti-
∙2007년~2017년 조선대학교 컴퓨터공
mation in basketball video.” Journal of Visual 학부 부교수
Communication and Image Representation, ∙2017년~현재 가천대학교 AI‧소프트웨어
학부 교수
Vol. 20(3), pp. 204-216, 2009 ∙관심분야: 컴퓨터비전, 패턴인식, 딥러닝,
[21] W. T. Chu, C. W. Wang, and J. L. Wu, 뇌과학, 생물정보학 등
“Extraction of baseball trajectory and physics-
based validation for single-view baseball
video sequences.”, 2006 IEEE International
Conference on Multimedia and Expo, pp.
1813-1816, IEEE, July, 2006

52

You might also like