Shot Detection from Discrete Location Data in Ice Hockey using Machine Learning
Ohtonen, Antti (2020)
2020
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2020-11-06
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202010067237
https://urn.fi/URN:NBN:fi:tuni-202010067237
Tiivistelmä
This thesis focused on comparing machine learning models for ice hockey shot detection using discrete time indoor location data of a puck. The three methods were an engineered model, a fully-connected neural network, and a convolutional neural network. The data used in the models was either samples of 100 consecutive data points of puck location data, x- and y-coordinates, or aggregate features derived from the samples – depending on the model.
The engineered model was a linear discriminant analysis algorithm trained with aggregate features derived from the puck location data. The features were selected by visual inspection of dimensionality reduced data. The dimensionality reduction was done by using uniform manifold approximation and projection algorithm. The features used were absolute values of x- and y-coordinates coupled with absolute values of x- and y-velocities.
The fully-connected and convolutional models used puck location data. Versions using the aggregate features were also tested. The model structure was selected by testing 20 different structures and selecting the one having the best performance in cross validation. The versions using aggregate features differed from the other models by the input dimensions only.
Cross validation results for all models were promising but results of evaluation with a manually annotated ice hockey game showed that the models are not usable in practice. The convolutional model using puck location had the highest accuracy in both tests, cross validation and evaluation. The accuracy of the fully-connected model using puck location was the second highest. The engineered model and the neural networks using aggregate features were not nearly as accurate.
The engineered model was a linear discriminant analysis algorithm trained with aggregate features derived from the puck location data. The features were selected by visual inspection of dimensionality reduced data. The dimensionality reduction was done by using uniform manifold approximation and projection algorithm. The features used were absolute values of x- and y-coordinates coupled with absolute values of x- and y-velocities.
The fully-connected and convolutional models used puck location data. Versions using the aggregate features were also tested. The model structure was selected by testing 20 different structures and selecting the one having the best performance in cross validation. The versions using aggregate features differed from the other models by the input dimensions only.
Cross validation results for all models were promising but results of evaluation with a manually annotated ice hockey game showed that the models are not usable in practice. The convolutional model using puck location had the highest accuracy in both tests, cross validation and evaluation. The accuracy of the fully-connected model using puck location was the second highest. The engineered model and the neural networks using aggregate features were not nearly as accurate.