Vertical Launch Angle Measurement of a Golf Ball Using Audio and

Monocular Video Data

Jan-Bernd Menke-Zumbr

agel

, Anna Brinkmann

and Andreas Hein

Assistance Systems and Medical Device Technology, Carl von Ossietzky University of Oldenburg, Germany

Keywords:

Golf, Launch Angle, Computer Vision, Sports.

Abstract:

In golf, several parameters can be measured that describe how the golf ball was hit and how the ball lifts after

impact with the golf club, the so-called launch parameters. In addition to the spin rate or the velocity of the ball

and club head, the launch angle is an important value that describes the vertical component of the ball’s launch

direction. For professional use, there are systems called launch monitors that use either radar-technology,

multiple high-speed cameras or a combination of both to measure the above parameters. Despite their high

accuracy, these systems can suffer from disadvantages such as some inconvenience regarding size, weight or

setup, and an inaccessibility regarding high cost. Therefore we present a method for vertical launch angle

measurement based on monocular low frame rate video and audio data, by detecting the motion blur structure

created by the launching golf ball. This approach allows the vertical launch angle to be measured with a

simple and inexpensive setup that achieves an accuracy of ± 0.74°, which is comparable to a commercial

launch monitor.

1 INTRODUCTION

In golf, there are different types of ball ﬂights used in

different situations on the golf course. They are de-

termined by different parameters and characteristics

of the golf swing and the way how the ball is hit. The

most important parameters are the launch parameters,

which describe how the ball lifts off after impact with

the golf club. The most common launch parameters

are listed and described in Table 1. With these param-

eters it is possible to calculate the outcome of the golf

swing regarding the trajectory of the golf ball.

While learning the golf swing or training for dif-

ferent ball trajectories a numerical feedback on these

launch parameters can be helpful to build consis-

tency and conﬁdence in different situations on the

golf course. In addition, certain combinations of

launch parameter values can be used to detect mis-

takes during the golf swing. The commercial so-

lutions for this type of ball ﬂight measurement are

called launch monitors. They come in a wide vari-

ety of price and accuracy levels, using photometric-

based systems, radar-based systems, or a combination

of both.

https://orcid.org/0009-0005-6412-0882

https://orcid.org/0000-0001-5228-4947

https://orcid.org/0000-0001-8846-2282

Figure 1: Visualization of a golf ball trajectory using a low

and high dynamic loft.

The radar-based systems, for example Trackman

4 (Trackman A/S, 023a) and Flightscope Mevo+

(FlightScope Mevo LLC, 2023), measure the move-

ment of the ball and club in three-dimensional

space, using the Doppler effect. By emitting radar

waves and measuring the reﬂection it is possible, for

example, to determine the speed of an object or the

ball spinning rate (Tuxen, 2010). This technology

allows to track a golf ball up to a distance of 365

meters (TrackMan A/S, 023b).

The photometric-based systems, for example GC-

Quad (Optimum Golf Technologies Ltd, 2023) and

Syktrak (GOLFTEC Enterprises LLC, 2023), are

using a set of high-speed cameras working at up to

120

Menke-Zumbrägel, J., Brinkmann, A. and Hein, A.

Vertical Launch Angle Measurement of a Golf Ball Using Audio and Monocular Video Data.

DOI: 10.5220/0012163200003587

In Proceedings of the 11th International Conference on Sport Sciences Research and Technology Support (icSPORTS 2023), pages 120-127

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

Table 1: Launch parameter overview (Johansson et al.,

2015).

Launch parameters Description

Club head speed

The velocity of the golf

club head at impact.

Face angle

The alignment of the

golf club face relative to

the target direction.

Swing path

The direction of the club

movement relative to the

target direction at im-

pact.

Angle of Attack

The vertical component

of the club movement at

impact.

Ball speed

The starting velocity of

the golf ball after impact.

Dynamic Loft

The vertical alignment of

the club face at impact.

Vertical

launch angle

Angle of the ball ﬂight

direction relative to the

horizon.

Horizontal

launch angle

Angle of the ball ﬂight

direction relative to the

target direction.

Spin Rate

The Rotation of the golf

ball per minute after im-

pact.

Spin Axis

The tilt of the spin axis

after impact.

10.000 frames per second, to capture the movement

of the ball and the club at impact from different

angles. By detecting the ball and its movements

in the ﬁrst few milliseconds of the ball ﬂight, it is

possible to calculate, for example, speed, spin rate or

direction (Kiraly, 2005).

One can divide the launch parameters into two

different categories: Firstly, the ball related param-

eters, such as launch direction, ball speed, spinning

rate. Secondly the club related parameters, such

as dynamic loft, angle of attack or club face angle.

(Leach et al., 2017) shows that in comparison with

an 3D motion capturing system, even though the

two mentioned measurement methods are completely

different, the measured ball related parameters of

both systems are highly consistent with each other.

One of the most important ball related parame-

ters is the launch direction, which can be represented

by horizontal and vertical components. While the

swing path and the club face angle are mainly

determining the horizontal component, the vertical

component is mainly determined by the angle of

attack and the dynamic loft at impact. (Wood et al.,

2018)

By adjusting the dynamic loft, different vertical

launch angles can be achieved. Referring to Figure

1 it can be seen, that by reducing the dynamic loft,

the vertical launch angle is also reduced. This, in

combination with a constant club head speed, results

in a signiﬁcantly lower ball ﬂight height and an

increased distance the ball rolls on the ground. On

the other hand, a higher launch angle results in a

higher ball trajectory and a shorter roll distance. Both

ﬂight trajectories can be useful in different situations

regarding obstacles on the golf course and are crucial

for achieving a low number of strokes.

When training to achieve a speciﬁc vertical launch

angle, a launch monitor can be helpful. Even though

the introduced commercial launch monitors can

achieve a high accuracy of ±0.3° using radar-based

systems and ±0.1° using photometric systems (Leach

et al., 2017), they suffer from certain disadvantages,

depending on the systems used. In addition to some

inconvenience due to size, weight, or setup, most

launch monitors require an additional device, such as

a smartphone or tablet, to display the measurements.

Also, the high cost (up to ∼20.000 $) doesn’t make

them suitable for amateur and beginning golf players.

Figure 2: Diagram of vertical launch angle measurement.

Therefore, we present a new method for vertical

launch angle measurement using only a monocular

low frame rate video in combination with the audio

data, as shown in Figure 2. The vertical launch angle

Vertical Launch Angle Measurement of a Golf Ball Using Audio and Monocular Video Data

121

measurement is performed by detecting the motion

blur created by the launching ball after impact.

Our approach leads to a simple and convenient

set up (shown in Figure 7), allowing to work with

only one device, such as a smartphone or tablet.

Additionally, the video data of the golf swing can

be recorded for further analysis, which has been

shown to have a positive impact on learning progress

(Guadagnoli et al., 2002).

2 METHODS

2.1 Swing Detection

To detect the launch angle of a golf ball after club

head impact from video data, the ﬁrst task is to detect

in which frame of the video the impact occurred. In

contrast to a method based on ball or club tracking

(Chotimanus et al., 2012; Zhang et al., 2020), where

either the golf ball or the club is observed to detect a

swing, we developed a method using the combination

of video and audio data of a recorded swing sequence

to detect an impact.

Figure 3: Audio data of a recorded training sequence in time

domain.

Looking at the audio data in the time domain

(Figure 3), we can detect peaks that indicate that an

event has occurred, such as the impact. The detection

of a peak is done by searching for a maximum value

between a rising and falling ﬂank, using a threshold

value calculated according to (1), where x(t) is the

audio signal in the time domain.

Using the peak time t

event

and the video data

frame rate f ps we can calculate the exact frame f

event

including the event, which created the mentioned

peaks in the audio signal, according to (2).

threshold = max(x(t)) ∗ 0.25 (1)

Figure 4: Visualization of ball detection: a) 3 frames be-

fore impact frame , b) 3 frames after impact frame, c)

grayscale image of the absolute difference between a) and

b), d) cropped and binarized image of golf ball.

event

= ⌊t

event

∗ f ps⌋ (2)

However, the detected peaks can also be caused by

different events, such as a practice swing where only

the swing movement was done without the intention

of hitting the ball, while the sound event was caused

by hitting the ground, an impact of a different player

or other non-speciﬁc sounds.

To increase the conﬁdence that a real swing oc-

curred at the audio-based detected moment of impact,

we further examine whether a golf ball left the scene

at the detected frame. Therefore, we subtract the

frame a speciﬁc time before and after the detected

impact frame from another as shown in Figure 4 a)-c).

To reduce the probability of occlusions in the

assumed area where the golf ball is located caused by

the club itself, the frame difference from the impact

frame at 60 fps is set to 3. That is, if an impact

is detected in frame f

event

, we subtract the frame

event

− 3 with the frame f

event

+ 3.

In combination with the assumption that the golf

ball is in the lower third of the image, it is possible

to detect a ball with a particularly high conﬁdence

using established computer vision methods. By

binarizing the image as shown in Figure 4 c) using

a threshold obtained from the Otsu method (Otsu,

1979), a golf ball leads to an explicit pixel structure

with a certain roundness and size. Using the Blob

Detector from OpenCV (OpenCV, 2023b) on the

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

122

Figure 5: Visualization of threshold impact: a) threshold =

10 , b) threshold = 20 c) threshold = 30, d) threshold = 40.

resulting image in Figure 4 d), we get the information

if a ball is included in the image and further the size

and position in pixel values. If the detection of a golf

ball is successful, the impact detection is marked as

valid, if there is no golf ball included in the resulting

image, the detection is marked as false.

In contrast to ball- or club-tracking approach,

using audio-event-triggered swing detection, we

achieve lower computing time and a independence

from possible occlusions of balls, when, for example,

setting up for a swing or other arbitrary movements

in the scene.

2.2 Launch Angle Measurement

With the resulting information from the swing detec-

tion as described above, namely the exact moment

when a swing occurred and the ball position in pixel

values, we address the problem of measuring the

launch angle of the golf ball using monocular video

data. Due to the high velocities that a golf ball can

reach while playing golf, averaging up to ∼270 km/h

(Hahn, Christian, 2017), an exact tracking of the golf

ball during ﬂight requires either a high-speed camera

or other technologies such as those mentioned in

the Introduction section. To solve this problem with

a low frame rate camera (60-120 fps), we use the

motion blur created by the launching ball in the

detected impact frame. The basis for this is a high

contrast between the golf ball and the background,

which is given by the golf sport itself, using mostly

white or yellow golf balls, which are hit either from a

green practice mat or grass.

To detect the motion blur caused by the ball,

we use a method similar to the one we used for swing

detection. Given the impact frame f

impact

, we created

two images f

−

and f

by calculating the absolute

difference between the image frame f

impact

and frame

impact

− 1 and also between frame f

impact

and frame

impact

+ 1. Subsequently, by calculating the addition

of these two resulting images f

−

and f

, we get the

resulting image where we perform the ball launch

detection on. In contrast to the previous method, it

is necessary to consider three images due to the low

frame rate combined with a lack of synchronization

between the moment of impact and the moment an

image is recorded. This could lead either to a small

motion blur structure in image f

−

or to the situation

that the ball already left the scene in image f

Figure 6: Processing steps: a) image subtraction and addi-

tion, b) result of a), c) preprocessed image of b) d) result of

line detection with ROI.

Combining the information included in the images f

−

and f

, by adding these two images to one another,

we increase the probability of detecting the desired

motion line, as seen in Figure 5 c).

To separate the motion blur caused by the club from

the motion blur caused by the golf ball, we binarize

the image with a certain threshold. Looking at each

pixel of a grayscale image, we compare the pixel’s

value with the before mentioned threshold. If the

Vertical Launch Angle Measurement of a Golf Ball Using Audio and Monocular Video Data

123

Data: Binary Image I

, Ball Position P

Result: Angle δ

Apply HoughLinesP from OpenCV and

assign results as list of lines L;

Compute ROI using Ball Position P

and

max. angle α =50°;

forall line in L do

if line in ROI then

Compute midpoint m and angle β of

line;

Compute angle γ between m and P

;

if D(β,γ) > 3° then

remove line

end

else

remove line;

end

Apply RANSAC Algorithm from SciPy on

resulting list of lines L;

Compute angle δ of resulting line;

Algorithm 1: Launch angle detection.

value is below the threshold, the pixel value is set

to 0 (black) and if the value is above the threshold,

the pixel value is set to 255 (white). Figure 5 shows

the effect of different threshold values while creating

a binary image. We can observe that by setting a

threshold to low (<20), the motion blur created by

the club overlays the structures created by the ball

(Figure 5 a)), on the other hand if the threshold is too

high (>30) it occurs, that the motion line created by

the launching ball is erased from the image (Figure

5 d)). The optimal threshold value is observed to be

between 20 and 30, as seen in Figure 5 b) and c), the

line created by the motion blur of the ball is clearly

visible and can be extracted from the image using

computer vision methods.

Algorithm 1 describes our approach for detect-

ing the launch angle, based on the binary image I

using computer vision methods. First, we specify

a region of interest (ROI), which is deﬁned by the

ball position achieved from the swing detection (2.1)

and an assumed minimum and maximum launch

angle, which is shown in Figure 6 d). To improve

the line detection using the probabilistic hough line

transform from OpenCV (OpenCV, 2023a), we apply

a sequence of computer vision methods including a

morphological opening, a horizontal Sobel ﬁlter and

a Canny Edge Detection (6 c)). The following line

detection returns a set of lines, of different length and

direction. By taking into account the minimum and

maximum launch angle condition, and additionally

the constraint that an extension of the lines must pass

the ball’s center position with a deﬁned offset, we

can perform a selection of lines that represent the

golf ball’s line of motion with a high probability.

To detect possible outlier, we ran the Random

Sample Consensus (RANSAC) algorithm (Fischler

and Bolles, 1981) for linear regression using the

endpoints of the selected lines.

The resulting line is considered to represent the

launch of the golf ball, with respect to the x-axis of

the image.

2.3 Experiment

For the evaluation of the presented launch angle mea-

surement approach, we use a radar-based launch mon-

itor, as described in the Introduction section, the

Garmin Approach R10 (Garmin Ltd. , Swiss, ∼600

$) (Garmin Ltd., 2023). The setup for this is shown in

Figure 7, where the camera used for our approach is

placed in front of the golf player at a distance of ∼2

meters from the golfer. The commercial launch mon-

itor is placed to the left of the golf player at a distance

of 2.1 meters from the ball.

Figure 7: Setup for evaluation.

To compare our approach with the results of the

launch monitor, we recorded 41 golf swings with two

different frame rates, 60fps and 120fps, while using

two different clubs, varying in loft angle and length.

The clubs used are a high loft pitching wedge (PW)

and a relative low loft 7 iron. The speciﬁcations are

listed in Table 2.

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

124

Table 2: Club speciﬁcation of ping rapture V2.

Club Loft angle [°] Length [m]

PW 45° 0.94 m

7 iron 32° 0.97 m

3 RESULTS & DISCUSSION

As the use of audio and monocular video data to esti-

mate golf launch parameters are underrepresented in

the literature, the evaluation of our method is cur-

rently limited to a comparison with the results ob-

tained using a launch monitor as described in section

2.3.

Figure 8: Deviation between the launch monitor results and

our approach using a framerate of 60 fps.

Figure 9: Deviation between the launch monitor results and

our approach using a framerate of 120 fps.

The results of the comparison for each swing

examined at either 60 fps or 120 fps are shown

in Figure 8 and 9 as the deviation between our

approach and the used launch monitor in degrees.

The tolerance of the Garmin Approach R10 regarding

the vertical launch angle measurement is given as ±

1°. The tolerance range is represented as a gray area

in Figure 8 and 9. The computation time for each

swing is observed to be ∼4 seconds on average with

a standard deviation of 1.3 seconds, using a 1,4 GHz

Quad-Core Intel Core i5 and an unoptimized code.

It can be seen that our approach provides results

that are in strong agreement with the measurements

of the Garmin Approach R10. Table 3 shows the

absolute average deviation for each used club at each

frame rate, along with the combined absolute average

deviation. In addition, it can be observed that the

overall absolute average deviation is smaller than the

tolerance range of the used ground truth system.

Figure 10: Absolute value comparison at 60 fps.

Figure 11: Absolute value comparison at 120 fps.

As we can see in Table 3, the deviation for the club

with the higher loft angle, the pitching wedge, is

signiﬁcant lower than the 7 iron. Due to the length

difference (see Table 2) between the clubs used,

the club head speed of the 7 iron is assumed to be

higher than the pitching wedge, which leads to the

assumption that the results are speed dependent. Fur-

thermore, isolated deviations can be observed, such

as swing No.25 (see Figure 8), which are assumed to

be related to the horizontal launch angle as described

in the following section. However, looking at the

absolute difference in Figure 10 and 11, we can see a

high correlation between the results of the different

measurement methods.

Vertical Launch Angle Measurement of a Golf Ball Using Audio and Monocular Video Data

125

Table 3: Absolute average deviation of our approach.

Club 60 fps 120 fps Combined

PW 0.63° 0.59° 0.61°

7 iron 0.77° 0.95° 0.86°

Combined 0.7° 0.77° 0.74°

4 LIMITS

The use of a monocular camera system results in lim-

itations regarding the accuracy of the measurement.

Due to the projection on a two-dimensional plane,

we have a loss of information about the horizontal

launch direction. If the horizontal launch angle is sig-

niﬁcantly higher or lower than 0°, meaning that the

launch direction is to the left or right of the target di-

rection, it will lead to a incorrect measurement. Fig-

ure 12 shows that if the launch direction is pointing

away from the camera plane, it will lead to a pro-

jection on a two-dimensional plane that indicates a

higher vertical launch angle than it actually is. Con-

sequently, it will be measured lower if the launch di-

rection is pointing towards the camera plane.

Figure 12: Perspective projection of different horizontal

launch angles with constant vertical launch angle.

Therefore, our approach only provides accurate

results when the horizontal launch angle is close to

zero, meaning the launch direction is similar to the

target direction or parallel to the camera plane.

5 CONCLUSION

We presented an approach to measure the vertical

launch angle of a golf ball using only video and audio

data in combination with a low frame rate, reaching

an accuracy equal to the results of a commercial

launch monitor, the Garmin Approach R10, using

a radar-based method. This means that with our

approach, a golf player has the ability to measure the

vertical launch angle of their golf shot without the

need for speciﬁc hardware such as a launch monitor,

just a camera and a PC, or any device that combines

the two such as a smartphone or tablet.

It should be noted that this approach is designed

for ofﬂine use only, meaning that we only analyzed

pre-recorded video data. However, looking at the

results in terms of computation time, our approach

has the potential to be an online application leading

to a system which is capable of providing direct

feedback to the golf player after the golf swing.

Therefore, only the audio event detection of the

swing detection algorithm needs to be adapted to

detect the club impact with the ball without the

information after the swing, for example by training

a CNN-based audio classiﬁer.

In addition to an online capability, the motion

blur detection will also be tested regarding club

related parameters such as dynamic loft, angle of

attack or club head speed. Concluding a detailed

data fusion of audio and video data has a great po-

tential to achieve a higher accuracy of our approach

and to develop new methods for launch parameter

measurements or golf assistance systems.

REFERENCES

Chotimanus, P., Cooharojananone, N., and Phimoltares, S.

(2012). Real swing extraction for video indexing in

golf practice video. In 2012 Computing, Communi-

cations and Applications Conference, pages 420–425.

IEEE.

Fischler, M. A. and Bolles, R. C. (1981). Random sample

consensus: a paradigm for model ﬁtting with appli-

cations to image analysis and automated cartography.

Communications of the ACM, 24(6):381–395.

FlightScope Mevo LLC (2023). Flightscope mevo -

portable golf launch monitors and simulators. https://

www.ﬂightscopemevo.eu/. Last checked on April 27,

2023.

Garmin Ltd. (2023). Garmin approach r10. https://

www.garmin.com/de-DE/p/695391/. Last checked on

April 27, 2023.

GOLFTEC Enterprises LLC (2023). The skytrak expe-

rience. https://www.skytrakgolf.com/launch-monitor.

Last checked on April 27, 2023.

Guadagnoli, M., Holcomb, W., and Davis, M. (2002). The

efﬁcacy of video feedback for learning the golf swing.

Journal of sports sciences, 20(8):615–622.

Hahn, Christian (2017). 2017 pga and lpga tour avg. https:

//blog.trackmangolf.com/2017-pga-lpga-tour-avg/.

Last checked on April 27, 2023.

Johansson, U., K

onig, R., Brattberg, P., Dahlbom, A., and

Riveiro, M. (2015). Mining trackman golf data. In

2015 International Conference on Computational Sci-

ence and Computational Intelligence (CSCI), pages

380–385. IEEE.

Kiraly, C. (EP 1 509 781 B1, Mar. 2005). Flight parameter

measurement system.

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

126

Leach, R. J., Forrester, S. E., Mears, A., and Roberts, J. R.

(2017). How valid and accurate are measurements of

golf impact parameters obtained using commercially

available radar and stereoscopic optical launch moni-

tors? Measurement, 112:125–136.

OpenCV (2023a). Hough line transform. https://docs.

opencv.org/3.4/d9/db0/tutorial hough lines.html.

Last checked on April 27, 2023.

OpenCV (2023b). Simpleblobdetector class refer-

ence. https://docs.opencv.org/3.4/d0/d7a/classcv 1

1SimpleBlobDetector.html. Last checked on April 27,

2023.

Optimum Golf Technologies Ltd (2023). Gcquad launch

monitor: Foresight sports. https://foresightsports.eu/

launch-monitors/gcquad/. Last checked on April 27,

2023.

Otsu, N. (1979). A threshold selection method from gray-

level histograms. IEEE transactions on systems, man,

and cybernetics, 9(1):62–66.

Trackman A/S (2023a). The complete performance solu-

tion. https://www.trackman.com/. Last checked on

April 27, 2023.

TrackMan A/S (2023b). What we track. https://www.

trackman.com/golf/trackman-4/what-we-track. Last

checked on April 27, 2023.

Tuxen, F. (EP2218483A2, Aug. 2010). Determination of

spin parameters of a sports ball.

Wood, P., Henrikson, E., and Broadie, C. (2018). The in-

ﬂuence of face angle and club path on the resultant

launch angle of a golf ball. In Proceedings, volume 2,

page 249. MDPI.

Zhang, T., Zhang, X., Yang, Y., Wang, Z., and Wang, G.

(2020). Efﬁcient golf ball detection and tracking based

on convolutional neural networks and kalman ﬁlter.

arXiv preprint arXiv:2012.09393.

Vertical Launch Angle Measurement of a Golf Ball Using Audio and Monocular Video Data

127