02/2021 – 08/2021
Monitoring the actions of a player in sports helps to prevent injuries from overuse or incorrect
techniques and can help to place emphasis on exercises targeting a specific movement [1, 2]. The
monitoring can either be done manually by watching a video recording of the event which is
time consuming and needs an expert of the field or by using a human activity recognition (HAR)
Human Activity Recognition (HAR) is the automated recognition and classication of activities
from a continuous stream of input sensors . Numerous papers have dealt with HAR in sports.
Most of them tried to classify sport specific movements or give feedback on the execution, like
Anand et al. did with analysing strokes in swing sports , Brock et al. with giving feedback on
style errors of ski jumpers  or Kasiri et al. with classifying boxing punches . Cust et al. provide
an overview of this research . HAR systems use different kinds of input signals. Mostly these
are image sequences (videos), sensor data or a combination of both. Inertial measurement units
(IMU) have proven to work well in sensor based HAR. IMUs typically contain accelerometers,
magnetometers or gyroscopes which measure continuously along the three axis of space. In the
stream of the sensor data HAR systems can detect when an action occurs and typically use
machine learning algorithms to try to classify the action .
In a paper from 2017 Kautz et al. collected data from volleyball players performing sport specific
movements by mounting an IMU at the wrist of the dominant hand of each player and performed a
recognition and classification of these movements. The authors discussed several algorithms based
on feature selection like a Naïve Bayes Classifier or support vector machines. All these algorithms
are outperformed by a Deep Convolutional Neural Network (DCNN) .
For any machine learning or deep learning algorithm to perform well a suficient set of training
data is needed. Huge sets of labeled data are generally either very costly to obtain or not available
at all. In order to overcome this limit Transfer Learning is used. The idea behind this approach
is to project already acquired knowledge to a new domain which not only reduces the time it takes
to train the classifier but also reduces the required size of the training dataset . There has been
some research which successfully used transfer learning in HAR and proved that neural networks
can be trained using a reduced number of training samples .
The goal of this thesis is to create a deep neural network which is based on the architecture
of the network by Kautz et al. and transfer the knowledge to the classication of typical types of
throws which occur in an ultimate frisbee match. The seven types of throws are: forehand inwards
moving throw, forehand outwards moving throw, forehand straight throw, backhand inwards moving
throw, backhand outwards moving throw, backhand straight throw and a hammer (a throw
over the top of your head). Furthermore a deep neural network which will only be trained on the
frisbee data is going to be created. The performance of this network will then be compared with
the earlier mentioned, pretrained network.
 Kautz, T., Groh, B.H., Hannink, J. et al.: Activity recognition in beach volleyball using a
Deep Convolutional Neural Network. Data Min Knowl Disc 31, 16781705 (2017).
 Schuldhaus, Dominik: Human Activity Recognition in Daily Life and Sports Using Inertial
Sensors. FAU University Press (2019)
 Anand, A., Sharma, M., Srivastava, R., Kaligounder, L., & Prakash, D.: Wearable motion
sensor based analysis of swing sports. In 2017 16th IEEE International Conference on
Machine Learning and Applications (ICMLA) (pp. 261267). (2017)
 Brock, H., Ohgi, Y., & Lee, J.: Learning to judge like a human: Convolutional networks for
classification of ski jumping errors. Proceedings of the 2017 ACM International Symposium
on Wearable Computers – ISWC ’17, 106113. (2017)
 Kasiri-Bidhendi, S., Fookes, C., Morgan, S., Martin, D. T., & Sridharan, S.: Combat sports
analytics: Boxing punch classication using over- head depth imagery. In 2015 IEEE International
Conference on Image Processing (ICIP) (pp. 45454549). Quebec City, Canada
 Emily E Cust, Alice J Sweeting, Kevin Ball & Sam Robertson: Machine and deep learning
for sport-specic movement recognition: a systematic review of model development and
performance. Journal of Sports Sciences, 37:5 (2019)
 Cook, D., Feuz, K.D. & Krishnan, N.C.: Transfer learning for activity recognition: a survey.
Knowl Inf Syst 36, 537556 (2013).
 Ramasamy Ramamurthy, S, Roy, N.: Recent trends in machine learning for human activity
recognition A survey. WIREs Data Mining Knowl Discov. 2018; 8:e1254. (2018)