Sensor-based stroke detection and stroke type classification in table tennis

Abstract

In this paper we present a sensor-based table tennis stroke detection and classification system. We attached inertial sensors to table tennis rackets and collected data of 8 different basic stroke types from 10 amateur and professional players. Firstly, single strokes were detected by a event detection algorithm. Secondly, features were computed and used as input for stroke type classification. Multiple classifiers were compared regarding classification rates and computational effort. The overall sensitivity of the stroke detection was 95.7% and the best classifier reached a classification rate of 96.7%. Therefore, our presented approach is able to detect table tennis strokes in time-series data and to classify each stroke into correct stroke type categories. The system has the potential to be implemented as an embedded real-time application for other racket sports, to analyze training exercises and competitions, to present match statistics or to support the athletes' training progress. To our knowledge, this is the first paper that addresses a wearable support system for table tennis, and our future work aims at using the presented results to build a complete match analysis system for this sport.

Full text

Sensor-based Stroke Detection and Stroke Type
Classification in Table Tennis
Peter Blank, Julian Hoßbach, Dominik Schuldhaus and Bjoern M. Eskofier
Digital Sports, Pattern Recognition Lab, Department of Computer Science
Friedrich-Alexander University Erlangen-N¨
urnberg (FAU), Erlangen, Germany
Corresponding author: peter.blank@cs.fau.de
ABSTRACT
In this paper we present a sensor-based table tennis stroke de-
tection and classification system with motion data of inertial
measurement units. We attached sensors to the front-end of
table tennis rackets to collect data of 8 different basic stroke
types from 10 amateur and professional players in training ex-
ercises. Firstly, single strokes were detected by a robust peak
detection algorithm out of the computed acceleration energy
signal. Secondly, sets of statistical features, correlation-based
features and signal characteristics were computed within a
specified data interval of each detected peak and used as in-
put for the stroke type classification. Multiple classifiers were
compared regarding classification rates and computational ef-
fort. The overall sensitivity of the stroke detection was 95.7%
and the best classifier reached a classification rate of 96.7%.
Therefore, our presented approach is able to detect table ten-
nis strokes in time-series data and can classify each stroke
into correct stroke type categories with satisfactory results.
The system has the capability to be implemented as an em-
bedded real-time application for wearable devices, to analyze
training exercises and competitions, to present match statis-
tics or to support the athletes’ training progress.
Author Keywords
inertial sensors; stroke detection and classification;
movement analysis; table tennis
ACM Classification Keywords
I.5.4 Pattern Recognition: Applications: Signal processing
INTRODUCTION
Recent advances in microelectronics, sensor integration and
data analysis techniques opened new possibilities for wear-
able technology and found its way into health, fitness and
sports applications [17]. While size and power consumption
of sensors and electronics made it difficult to obtain physio-
logical and kinematic motion data in the past, today, miniatur-
ized circuits, powerful microcontrollers and energy-friendly
wireless data transmission allow Wearables and sensor de-
vices to be integrated into different body-worn accessories
Paste the appropriate copyright statement here. ACM now supports three different
copyright statements:
•ACM copyright: ACM holds the copyright on the work. This is the historical ap-
proach.
•License: The author(s) retain copyright, but ACM receives an exclusive publication
license.
•Open Access: The author(s) wish to pay for the work to be open access. The addi-
tional fee must be paid to ACM.
This text field is large enough to hold the appropriate release statement assuming it is
single spaced.
Every submission will be assigned their own unique DOI string to be included here.
such as garments, wristbands, shoes, smartphones or sports
equipment like balls or rackets.
If Wearables are used in a sports science context, inertial sen-
sor data based on accelerometer and gyroscope signals, is the
most important source for state-of-the-art motion and move-
ment analysis [14]. The analyzing and visualizing of sensor
data is not only restricted to support professional athletes, but
can also be used for educational applications, statistical in-
formation and entertainment purposes [3]. Sensor-based mo-
tion analysis enables also an objective view on different kinds
of sports with complex and highly dynamic motions. These
are often associated with a technically correct performance
of racket movements using additional sports equipment. In
[7] and [1], tennis players already benefit from sensor-based
stroke classification and swing investigation in terms of a
coaching tool and a feedback indicator. Other examples de-
scribe the professional evaluation of baseball pitchers and
batters using distributed wearable sensor systems [16] and
an instant feedback system based on an inertial sensor and a
smartphone for golf putting [13]. Its many possible strokes to
achieve spin and speed variations, definitely make table ten-
nis another interesting and promising candidate for research.
There are basic table tennis strokes like drive (perpendicular
hit without spin), push (hit below the ball, resulting in back-
spin), block (hit with less movements) or topspin (hit above
the ball, resulting in topspin), resulting in different kinds of
spin (backspin, topspin or sidespin). All stroke types can be
performed with the forehand and the backhand. Furthermore,
many variations are possible due to a huge amount of rubbers
and rackets. In addition to normal rubbers, frictionless rub-
bers and pimples are available, which massively change the
spin behavior. Moreover, different grip styles such as shake-
hand or pen-holder (mainly used in asia) have become ac-
cepted.
All these variations could be analyzed using sensor-based ap-
proaches. The purpose of our work was the first investigation
on an automated sensor-based stroke detection and classifica-
tion system using a single inertial sensor attached to the front-
end of a table tennis racket handle. Kinematic data based on
acceleration and angular velocity was collected to compute
stroke events and to classify basic stroke types using pattern
recognition algorithms.
So far, only a few publications on table tennis movement anal-
ysis exist. They usually deal either with video-based methods
or sensor-based approaches. In [6], the table tennis table, the
ball trajectories, as well as players’ movements and actions
are tracked using multiple camera-based data sets. After-

ward, results are provided to broadcast table tennis videos.
Other approaches like [21] use motion capturing systems for
tracking targets during high speed movements. Multiple op-
tical markers are placed around a table tennis racket to sim-
ulate the hand and arm movements of a player. These mo-
tions are then used as input for a virtual reality application.
A combination of triaxial accelerometer data and 2 camera-
based videos in [23] is used to identify forehand and back-
hand stroke motions as well as the location of the ball and
the player relative to the table. For quantitative analysis of
table tennis block strokes, a pilot study is presented by [10]
using a 3-dimensional accelerometer attached with the use of
a wristband. Several block strokes taken from elite and ama-
teur players are averaged and compared to allow conclusions
about their different skill levels. In [5] an inertial sensor is
placed inside the handle of a table tennis racket measuring
acceleration and angular velocity. Data from different stroke
types is also averaged and visually processed for quantitative
analysis.
However, both sensor-based approaches in [10] and [5] do
not address a methodology for automated stroke detection
and classification. In this contribution we extended existing
sensor-based motion measuring methods with pattern recog-
nition algorithms. Our research focused on an automated
stroke detection and classification system of multiple stroke
types from players with playing abilities ranging from ama-
teur to professional players.
METHODS
Data Acquisition
Hardware Equipment
We used the miPod sensor platform [4] to collect all inertial
data from each player. This platform includes among other
sensors a triaxial accelerometer and a triaxial gyroscope re-
sulting in a six-dimensional data set for each player. The ac-
celerometer range was set to ±16 g and the gyroscope rate
was set to ±2000 ◦/s. Data was obtained with an analog-
digital converter resolution of 16 bits. We chose a sampling
rate of 1000 Hz to capture all highly dynamic motions. Mea-
sured data was stored on the internal flash module and was
processed offline. The video reference for stroke labeling
was provided by a CASIOrExilim HS EX-ZR200 high
speed camera with a frame rate of 120 fps and a resolution
of 640 x480 pixels. The sensor was attached to the front end
of every player’s racket handle using a strong double-sided
adhesive tape. To avoid misalignments of the sensor position,
the sensor was parallel adjusted in Y-direction of the handle
edges. Figure 1 exemplarily shows the sensor position and
the corresponding axes.
Study Design
A research study was planned to collect data of stroke types
of 10 players. This study was comprised of 2 female and 8
male players within an age between 26 and 50 years. All
were right-handed and used a shakehand grip. All partici-
pants were members of the German national table tennis as-
sociation. According to a continuously updated ranking co-
efficient they were split into 3 higher level and 7 lower level
Figure 1. miPod sensor attached to the front-end of the racket handle.
The coordinate system is aligned to the main axes of the racket: X-axis
and Y-axis span the main movement plain of the most common table
tennis strokes. Changes in spin become distinct in the direction of the
Y-axis. The Z-axis has most impact on differentiation between forehand
and backhand strokes.
players. We intentionally asked subjects of different play-
ing abilities to receive a high variability in the stroke pat-
terns. The 4 basic stroke types drive,push,block and topspin
were performed both with forehand and backhand. Overall,
we collected data for 8 different stroke types of each player.
The data collection is based on a two-player exercise session
divided in 8 one-minute long sub-exercises in a fixed order.
First, both player perform a forehand drive, then a forehand
push, followed by a forehand topspin of player A in combina-
tion with a forehand block of player B and a forehand topspin
of player B in combination with a forehand block of player A.
Afterward, these 4 sub-exercises are repeated with the back-
hand. The complete process was recorded on video (Figure
2 presents some sub-exercises) as well as documented in a
study protocol. An overview of the exercise session is shown
in Table 1.
No. Player A Player B
1 (FD) Forehand drive (FD) Forehand drive
2 (FP) Forehand push (FP) Forehand push
3 (FB) Forehand block (FT) Forehand topspin
4 (FT) Forehand topspin (FB) Forehand block
5 (BD) Backhand drive (BD) Backhand drive
6 (BP) Backhand push (BP) Backhand push
7 (BB) Backhand block (BT) Backhand topspin
←−−−−−−−−−−−−−
8 (BT) Backhand topspin (BB) Backhand block
Table 1. Session protocol of a two-player exercise including all one-
minute sub-exercises (No. 1 to 8): forehand drive (FD), forehand push
(FD) forehand topspin (FT), forehand block (FB), backhand drive (BD),
backhand push (BP), backhand topspin (BT) and backhand block (BB).
Classification System
The processing pipeline of our proposed classification system
consists of two parts. Firstly, strokes were detected using pre-
processing methods and a robust peak detection. Secondly,

Figure 2. Example sub-exercises: figure (a) shows a rally of forehand
topspins (FT) of left player and forehand blocks (FB) of right player,
figure (b) shows a rally of backhand topspins (BT) of left player and
backhand blocks (BB) of right player, figure (c) shows a rally of forehand
pushs (FP) and figure (d) shows a rally of backhand pushs.
the extracted strokes were used to generate feature sets for
the stroke type classification system. The complete process is
summarized in Figure 3. This dual-tracked approach prese-
lects specific events of interests for the stroke type classifica-
tion and finally reduces the amount of data.
Figure 3. Processing pipeline for stroke detection (energy calculation us-
ing acceleration data, high-pass filtering, negative value masking, peak
detection) and classification (stroke interval computation using accelera-
tion data and angular velocities, feature extraction, normalization, clas-
sification).
Stroke detection
The stroke detection is based on the accelerometer signals
of the X-axis, the Y-axis and the Z-axis. Firstly, the signal
energy
Eraw (xi) = ax(xi)2+ay(xi)2+az(xi)2(1)
of all accelerometer axes was calculated (1). It is a significant
indicator for activity recognition [22]. Secondly, a high-pass
Butterworth filter with an order of n= 1 was used to sup-
press arm movements and to emphasize rapid changes in the
computed energy signal, which indicate ball-racket impact
events. According to [20] the minimal duration of the arm
movements performing a fast table tennis stroke is less than
18 ms. Therefore, a cut-off frequency of fc= 25 Hz was cho-
sen. The filter was applied in a forward and backward manner
to avoid any filter delay. Next, all negative values were set to
zero (2), whereas positive values were kept unchanged.
S(xi) = Efilt(xi)i,if xi>0
0,otherwise (2)
Finally, a robust peak detection similar to [18] was per-
formed. It was based on a threshold th (3) and a segmentation
range k(4).
th =m+h·σ h ∈R+
0(3)
k= [xi−a;xi+b](4)
The threshold value th depends on a combination of the mean
mand the standard deviation σ. The parameters aand bof the
segmentation range roughly define the estimated peak width
in samples, which corresponds to the stroke duration in mil-
liseconds. A peak (5) was detected if the maximum value
within the segmentation range kexceeded the threshold th

and if the peak was not small in global context.
P(xi) = (1,if (max(S(xi)k>th)−m)>(h·σ)
0,otherwise (5)
An overview of the stroke detection pipeline with example
data of consecutive forehand drives can be found in Figure 4.
Figure 4. Example peak detection pipeline of a consecutive 8-second
forehand drive (FD) sub-exercise section from one player: (a) raw ac-
celeration data of all 3 axes, (b) energy calculation, (c) high-pass But-
terworth filter with n= 1 and fc= 25 Hz, (d) negative masking and
peak detection with th = 8.80 (blue dotted line), m= 0.44,σ= 4.18,
h= 2.2,a= 500 and b= 500. Blue stars represent detected peaks.
Stroke Classification
Based on the detected peaks, the stroke intervals were gen-
erated around the particular peak samples xpof the raw ac-
celerometer and gyroscope signals. Therefore, every stroke
interval was defined in the range [xp−α;xp+β]similar to the
peak segmentation range including all important components
of a single stroke. The stroke interval was set to 1000 ms, the
parameters αand βwere chosen to be 600 ms and 400 ms,
respectively. Since the countermovements before ball con-
tact reveal more meaningful information about the stroke type
[20], the interval window was shifted to past data. For ev-
ery player, sets of multiple data intervals for each stroke type
were computed. Subsequently, these data sets were provided
for the feature extraction step.
Features were calculated from all stroke intervals and all
axes. Generic time-domain features including statistical mo-
ments (mean, standard deviation, skewness, kurtosis) and sig-
nal characteristics (minimum, maximum, energy, median, in-
terquartile range) were computed. These have been proven
for activity recognition, event detection [2] and successful
classification of kinematic bio-signals [15]. Additionally,
heuristic inter-axis correlation features between each of the
axes of one sensor type were computed. All in all, 60 fea-
tures for every stroke interval were calculated and normalized
between [0; 1].
Classifiers were trained using the Embedded Classification
Toolbox (ECST) [19] and the WEKA data mining software
[11]. The ECST provides training of different classifica-
tion systems, evaluates classification rates and analyzes their
computational complexity. Since there is no best classi-
fier for a specific classification task [8], 6 different types of
classifiers were compared: Na¨
ıve Bayes (NB), RandomFor-
est (RF), Support Vector Machine (SVM) with linear kernel
(LIN) and radial based funtion (RBF) kernel, k-nearest neigh-
bours (kNN) and PART (PT) [9] as a rule based classifier. A
grid search was performed to optimize the cost parameter c∈
{1,10,100,1000}with logarithmic steps of the SVM [12].
RF and kNN were analyzed for ntrees ∈ {10,25,50,100}
and k∈ {1,2,3,4,5}, respectively.
Evaluation
The evaluation was done separately for stroke detection anal-
ysis and stroke type classification.
For stroke detection, all attempts of strokes were manually
labeled by a table tennis expert with a long-term experience.
These include strokes were no ball was hit, strokes which
should not have been part of the appropriate sub-exercise,
failed strokes, services and other random racket movements
with impact characteristics. The evaluation compared all de-
tected strokes with all labeled strokes of each player exercise-
wise. Matches were marked as true positives (#TP). Detected
strokes were marked as false positives (#FP) if they occurred
at positions where no labeled strokes were found. Undetected
labeled strokes were marked as true negatives (#TN), respec-
tively. Since there was no discrete window-based processing,
it is impossible to state false negatives (#FN).
Precision =#T P
#T P + #F P (6)
Recall =#T P
#T P + #F N (7)
F-Measure = 2 ·Precision ·Recall
Precision +Recall (8)
Therefore, the stroke detection algorithm was qualitatively
described with Precision (6), Recall (7) and the combined F-
Measure (8) as harmonic mean. Additionally, different peak

detection thresholds based on combinations out of mean m,
standard deviation σand the factor hwere evaluated.
The stroke type classification was based on all #TP and there-
fore contained only valid stroke categories. For the perfor-
mance assessment of all classifiers, the mean and class de-
pendent accuracy as well as mean overall classification rates
were computed with a 10-fold-cross validation. Furthermore,
algorithm complexity was estimated in order to obtain the
computational effort.
RESULTS
Stroke Detection
The sums of all labeled strokes from all players are shown in
Table 2. In total, 3004 strokes were labeled, whereof 1982
Type Labeled valid strokes Detected valid strokes
FD 404 404
FP 187 183
FB 250 249
FT 297 296
BD 278 277
BP 181 180
BB 161 161
BT 224 221
Total 1982 / 3004 1971 / 3097
Table 2. List of all labeled valid and detected valid strokes divided into
each stroke type category. Some strokes were executed with a higher
certainty than other strokes. Overall, 3004 strokes were labeled, whereof
1982 were valid strokes.
strokes counted as valid for a all sub-exercises. In contrast,
the peak detection algorithm counted a total of 3097 strokes,
whereof 1971 were marked as true positives (#TP). These
data sets were used as input for the subsequent stroke type
classification. The peak detection threshold results are given
in Figure 5 and Table 3. Different threshold combinations
th =m+h·σvarying h= [0; 5] in steps of 0.1 result in
a maximum F-Measure value of 96.9% at h= 2.2. This in-
duced a mean Precision and a Recall for all player-dependent
thresholds of 95.7% and 98.2%, respectively. Quantitative
Figure 5. Threshold dependent F-Measure values (gray) of all players
with mean (black). Varying thresholds th =m+h·σwith h= [0; 5] in
steps of 0.1result in a maximum mean F-Measure of 96.9% at h= 2.2.
results of extracted stroke types were presented by plotting all
consecutive valid strokes of one category. The patterns were
supplied with mean (black line) and standard deviation (dot-
ted lines) to illustrate stroke variances and differences in the
signal patterns. The interval window length was set to 1000
Threshold Precision Recall F-Measure
th =m+ 2.2·σ0.957 0.982 0.969
Table 3. Threshold evaluation results of the stroke detection based on
the F-Measure. Best result holds a Precision of 95.7% and a Recall of
98.2%.
Figure 6. Consecutive forehand stroke data with mean mand standard
deviation σin gyroscope X-direction: forehand drive (FD), forehand
push (FP), forehand block (FB) and forehand topspin (FT).
milliseconds, whereof 600 milliseconds before and 400 mil-
liseconds after ball contact were presented. Figure 6 and 7
show example gyroscope data in X-direction and Y-direction
of forehand strokes, whereas Figure 8 and Figure 9 repre-
sent example gyroscope data in X-direction and Y-direction
of backhand strokes. These consecutive stroke patterns were
collected of one single player during a complete exercise.
Stroke Type Classification
Table 4 shows the overall mean classification rates and the
computational efforts of all classifiers. The classifier with the
best performance was Support Vector Machine (SVM) with
the radial based function (RBF) kernel and a cost parameter
c= 10 with an overall classification rate of 96.7%. Table 5
presents the mean class-dependent classification rates of the
best classifier regarding all 8 stroke type categories. Best de-
tected stroke categories were forehand push with an accuracy
of 100.0% and backhand push with an accuracy of 99.4%.
Table 6 shows the confusion matrix of the proposed classi-
fication system. Labeled strokes and computed strokes are
compared with the corresponding rows and columns. Each
field reflects the class-dependent decision.
DISCUSSION
The results show that it is possible to detect table tennis
strokes and to classify different types of strokes with high

Figure 7. Consecutive forehand stroke data with mean mand standard
deviation σin gyroscope Y-direction: forehand drive (FD), forehand
push (FP), forehand block (FB) and forehand topspin (FT).
Figure 8. Consecutive backhand stroke data with mean mand standard
deviation σin gyroscope X-direction: backhand drive (BD), backhand
push (BP), backhand block (BB) and backhand topspin (BT).
accuracy. Overall, we collected 3004 strokes of 10 players
within 8 different sub-exercises (Table 2). However, only
1982 strokes were labeled as valid strokes. The reason why
nearly one third of all strokes could not be used for processing
is that services, netballs, edgeballs and wrongly performed
Figure 9. Consecutive backhand stroke data with mean mand standard
deviation σin gyroscope Y-direction: backhand drive (BD), backhand
push (BP), backhand block (BB) and backhand topspin (BT).
SVM
Type NB RF LIN RBF kNN PT
Accuracy
[%] 87.1 95.7 95.6 96.7 94.7 89.0
Effort mid lo mid hi hi lo
+,−472 n/a 1680 87584 114321 0
∗,÷2832 n/a 1652 46004 116289 0
ex,√x944 n/a 0 742 1971 0
≤35 n/a 7 35 5911 167
Table 4. Overall accuracy and computational effort (lo, mid, hi) regard-
ing number of executed operations of all classifiers. The RBF SVM clas-
sifier with cost parameter c= 10 was identified as best performing clas-
sifier with an overall accuracy of 96.7%. The linear SVM performed
best with c= 1, RF with ntrees = 100 and kNN with k= 3.
Stroke type Accuracy [%]
FD 99.3
FP 100.0
FB 95.2
FT 98.6
BD 94.2
BP 99.4
BB 87.6
BT 96.4
Table 5. Mean class-dependent classification rate of the best performing
classifier regarding the stroke type categories. Best detected stroke types
were forehand push (FP) and backhand push (BP) with a mean accuracy
of 100.0% and 99.4%, respectively.
strokes are included. This holds true both for amateur players
and professional players.

Type FD FP FB FT BD BP BB BT
FD 401 0210000
FP 0 183 000000
FB 12 0 237 00000
FT 4 0 0 292 0000
BD 0 0 0 0 261 0 14 2
BP 01000179 0 0
BB 0 1 0 0 19 0 141 0
BT 0000800213
Table 6. Confusion matrix of the best classification system. Rows repre-
sent labeled strokes, whereas columns indicate classified strokes.
The event detection method was able to detect strokes with a
Precision of 95.7 % and a Recall of 98.2 %. However, it has
to be considered, that these good results were achieved by a
study evaluation for the purpose of stroke classification. All
players were asked to perform pre-defined exercise sessions
with fixed numbers, durations and orders of sub-exercises.
During these sub-exercises, consecutive strokes with mostly
constant movement sequences were performed. Data of a real
match or a complex training session comprising of more than
one sub-exercise, would lead to the detection of more non-
stroke events (#FP). This might be caused by a higher influ-
ence of movement artifacts and inaccurate stroke exertions.
Nevertheless, a slightly lower Precision would not worsen the
final detection algorithm considerably.
The best classification results were achieved with the SVM
and the RBF kernel with an accuracy of 96.7%. Neverthe-
less, the high classification rate leads to a high computational
effort with a lot of numerical operations. As the final im-
plementation of the system is going to use simple micro-
controllers included in Wearables, it might be advantageous
to use classifiers with lower accuracy. This would result in
considerably less computational effort. Promising candidates
would be the linear SVM or the rule based PART algorithm.
Class-dependent classification rates can be seen in Table (5)
and misclassifications are shown in the confusion matrix (Ta-
ble 6). The highest number of incorrectly classified strokes
occurred for forehand block (FB) and backhand block (BB).
Both stroke types have similar signals compared to drive or
topspin, which can be quantitatively seen in the stroke pattern
figures (Figure 6 to Figure 9). Every player performed block
strokes in a more or less active way, which depends on the
style and ability of the player himself and the speed and spin
of the approaching ball. In contrast, the distinction of nearly
all forehand strokes from backhand strokes, as well as strokes
with backspin (FP, BP) from strokes with no or topspin (FD,
FB, FT, BD, BB, BT) showed no such problems. This is
caused due to a fundamental change in the stroke movement
(racket impact on the lower ball side or racket impact on the
upper ball side) resulting in a high Z-axis variation.
Non-valid strokes that were mistakenly detected were not
considered for the classification. Because of that high classifi-
cation rates were observed. Further methods have to circum-
vent this limitation to be able to detect events, which should
not be assigned to any of the specified stroke type categories.
One possible solution could be the implementation of a null-
class for non-stroke events. The consequence would proba-
bly be a lower classification rate, because non-stroke events
can have similar signal appearances as specified stroke types.
Services, for example, could be confused with forehand or
backhand drives. Therefore, a deeper look into the stroke in-
terval data becomes necessary to extract more stroke specific
information.
There has to be further focus on the real-time ability of the
system, in order to use it as a wearable application. This study
already showed first experiments of the computational effort
using different classifiers, but an advanced analysis of clas-
sification accuracy depending on computational effort would
be necessary. This also includes the investigation of lower
sampling rates and a window-based stroke detection for real-
time processing.
SUMMARY AND OUTLOOK
In this contribution we presented an approach for table ten-
nis stroke detection and stroke type classification using in-
ertial sensor data. We conducted a study with 10 subjects
and collected data for 8 basic table tennis stroke types in
a pre-defined exercise session. A stroke detection algo-
rithm consisting of energy calculation, high-pass filtering and
threshold-based peak detection was performed to detect all
strokes. We achieved an overall Precision of 95.7% and Re-
call of 98.2%. Based on the detected strokes, different fea-
tures were extracted and multiple classifiers were evaluated.
The best performance was achieved with a Support Vector
Machine (RBF kernel) yielding a classification rate of 96.7%.
With this study, we proved our system to be a reliable base for
future work in the area of sensor-based table tennis analyzing.
In the future, we will implement the whole processing
pipeline into embedded systems with less memory and com-
putational power. The goal is to develop a small wearable
device, which can be attached to the body using a wristband
or invisibly integrated into the table tennis racket. Thereby,
competition matches could be analyzed and arm and racket
movements classified to provide statistical information. Fur-
thermore, amateur players could use the system as a training
device to improve their own skills and to share their perfor-
mances with other dedicated players. Additionally, we will
try to extend the system to classify further strokes like chops
or strikes and to estimate the ball spin after the ball-racket
impact.
ACKNOWLEDGMENTS
We thank the table tennis club SpVgg Erlangen and all partic-
ipants for their support in this project. This work was further
supported by the Bavarian Ministry for Economic Affairs,
Infrastructure, Transport and Technology and the European
Fund for regional development.
REFERENCES
1. Ahmadi, A., Rowlands, D. D., and James, D. A.
Investigating the translational and rotational motion of
the swing using accelerometers for athlete skill
assessment. In Proceedings of the Conference on
Sensors (2006), pp. 980–983.

2. Avci, A., Bosch, S., Marin-Perianu, M., Marin-Perianu,
R., and Havinga, P. Activity recognition using inertial
sensing for healthcare, wellbeing and sports
applications: A survey. In Proceedings of International
Conference on Architecture of Computing Systems
(ARCS) (2010), pp. 1–10.
3. Baca, A., and Kornfeind, P. Rapid feedback systems for
elite sports training. Pervasive Computing vol. 5, no. 4
(2006), pp. 70–76.
4. Blank, P., Kugler, P., Schlarb, H., and Eskofier, B. M. A
wearable sensor system for sports and fitness
applications. In Proceedings of the Annual Congress of
the European College of Sports Sciences (ECSS) (2014),
pp. 703.
5. Boyer, E., Bevilacqua, F., Phal, F., and Hanneton, S.
Low-cost motion sensing of table tennis players for real
time feedback. International Journal of Table Tennis
Sciences vol. 8 (2013).
6. Chen, W., and Zhang, Y.-J. Tracking ball and players
with applications to highlight ranking of broadcasting
table tennis video. In Proceedings of the
Multiconference on Computational Engineering in
Systems Applications (IMACS), vol. 2 (2006), pp.
1896–1903.
7. Connaghan, D., Kelly, P., O’Connor, N. E., Gaffney, M.,
Walsh, M., and O’Mathuna, C. Multi-sensor
classification of tennis strokes. In Sensors Journal
(2011), pp. 1437–1440.
8. Duda, R. O., Hart, P. E., and Stork, D. G. Pattern
Classification. John Wiley & Sons, 2012.
9. Eibe, F., and Witten, I. H. Generating accurate rule sets
without global optimization. In Proceedings of the
International Conference on Machine Learning (ICML)
(1998), pp. 144–151.
10. Guo, Y.-W., Liu, G.-Z., Huang, B.-Y., Zhao, G.-R., Mei,
Z.-Y., and Wang, L. A pilot study on quantitative
analysis for table tennis block using a 3d accelerometer.
In Proceedings of International Conference on
Information Technology and Applications in
Biomedicine (ITAB) (2010), pp. 1–4.
11. Hall, M., Frank, E., Holmes, G., Pfahringer, B.,
Reutemann, P., and Witten, I. H. The weka data mining
software: an update. ACM SIGKDD explorations
newsletter vol. 11, no. 1 (2009), pp. 10–18.
12. Hsu, C.-W., Chang, C.-C., Lin, C.-J., et al. A practical
guide to support vector classification, 2003.
13. Jensen, U., Kugler, P., Dassler, F. A., and Eskofier, B. M.
Sensor-based instant golf putt feedback. Proceedings of
the International Symposium on Computer Science in
Sport (IACSS) (2011), pp. 49–53.
14. Jensen, U., Prade, F., and Eskofier, B. M. Classification
of kinematic swimming data with emphasis on resource
consumption. In Proceedings of the International
Conference on Body Sensor Networks (BSN) (2013), pp.
1–5.
15. Jensen, U., Ring, M., and Eskofier, B. M. Generic
features for biosignal classification. Sportinformatik
2012 (2012), pp. 162–168.
16. Lapinski, M., Berkson, E., Gill, T., Reinold, M., and
Paradiso, J. A. A distributed wearable, wireless sensor
system for evaluating professional baseball pitchers and
batters. In International Symposium on Wearable
Computers (ISWC) (2009), pp. 131–138.
17. McCann, J., and Bryson, D. Smart Clothes and
Wearable Technology. Elsevier, 2009.
18. Palshikar, G., et al. Simple algorithms for peak detection
in time-series. In In Proceedings of International
Conference of Advanced Data Analysis, Business
Analytics and Intelligence (IMAA) (2009).
19. Ring, M., Jensen, U., Kugler, P., and Eskofier, B.
Software-based performance and complexity analysis
for the design of embedded classification systems. In
Proceedings of the International Conference on Pattern
Recognition (ICPR) (2012), pp. 2266–2269.
20. Rodrigues, S. T., Vickers, J. N., and Williams, A. M.
Head, eye and arm coordination in table tennis. Journal
of Sports Sciences vol. 20, no. 3 (2002), pp. 187–200.
21. Rusdorf, S., and Brunnett, G. Real time tracking of high
speed movements in the context of a table tennis
application. In Proceedings of the ACM symposium on
Virtual Reality Software and Technology (2005), pp.
192–200.
22. Schuldhaus, D., Leutheuser, H., and Eskofier, B. M.
Towards big data for activity recognition: A novel
database fusion strategy. In Proceedings of the
International Conference on Body Area Networks (BSN)
(2014), pp. 97–103.
23. Sørensen, V., Ingvaldsen, R. P., and Whiting, H. T. A.
The application of co-ordination dynamics to the
analysis of discrete movements using table-tennis as a
paradigm skill. Biological Cybernetics vol. 85, no. 1
(2001), pp. 27–38.