Learning-Based Detection of Events in Eye-Tracking Data: An Investigation Into Small-Scale Models for Automotive Applications
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Engineering mathematics and computational science (MPENM), MSc
Publicerad
2024
Författare
Due, Martin
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
Detection of eye movement events in eye-tracking data is integral to various research
fields and commercial applications. Traditionally, the detection has been
accomplished either by hand or with threshold based detectors with the inherent
drawback that thresholds levels and other parameters had to be hand picked for
each scenario. In recent years machine learning methods have been employed for
eye movement event detection that do away with that requirement. However, these
models tend to be too large to run on limited resources in embedded applications,
particularly automotive ones. This thesis focuses on creating models that are smaller
but with retained performance. Five machine learning methods were evaluated and
hyperparameters were tuned to create well performing small models. Furthermore,
the usage of synthetic data for training was investigated, both as a supplement to
real data and as a sole source of training data. The study found that a Multilayer
Perceptron model (MLP) trained on a combination of real and synthetic data
struck the best balance between size and performance. Additionally, results show
that models trained purely on synthetic data also performed reasonably well. The
findings of this thesis suggest that small, efficient models can effectively detect eye
movement events, with potential applications in automotive contexts.
Beskrivning
Ämne/nyckelord
Eye movements, Event detection, Fixations, Saccades, Machine Learning, Model Optimization.