Author, Institution: Romas Vijeikis, Kaunas University of Technology
Science area, field of science: Technological Sciences, Informatics Engineering, T007
Research supervisor: Prof. Dr. Vidas Raudonis (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)
Dissertation Defence Board of Informatics Engineering Science Field:
Prof. Dr. Renaldas Urniežius (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007) – chairperson
Prof. Dr. Dalius Mažeika (Vilnius Gediminas Technical University, Technological Sciences, Informatics Engineering, T007)
Prof. Dr. Masaki Ogura (Hiroshima University, Japan, Technological Sciences, Informatics Engineering, T007)
Prof. Dr. Roel Pieters (Tampere University, Finland, Technological Sciences, Informatics Engineering, T007)
Prof. Dr. Algimantas Venčkauskas (Kaunas University of Technology, Technological Sciences, Informatics Engineering, T007)
Dissertation defence meeting will be at Rectorate Hall of Kaunas University of Technology (K. Donelaičio 73–402, Kaunas)
The doctoral dissertation and summary are available at the library of Kaunas University of Technology (Gedimino g. 50, Kaunas) and on the internet: R. Vijeikis el. dissertation.pdf
© R. Vijeikis, 2026 “The text of the thesis may not be copied, distributed, published, made public, including by making it publicly available on computer networks (Internet), reproduced in any form or by any means, including, but not limited to, electronic, mechanical or other means. Pursuant to Article 25(1) of the Law on Copyright and Related Rights of the Republic of Lithuania, a person with a disability who has difficulties in reading a document of a thesis published on the Internet, and insofar as this is justified by a particular disability, shall request that the document be made available in an alternative form by e-mail to doktorantura@ktu.lt.”
Annotation: The dissertation investigates video-based human behaviour recognition in two safety-critical domains: driver monitoring and violence detection. A unified clip-based methodological system is developed to evaluate spatio-temporal models with different computational characteristics under consistent conditions. Two compact student models are analysed: a hybrid 3D-ResNet-18+BiLSTM model and a PoI-ViT model composed in this work by combining a compact vision-transformer backbone with patch-of-interest selection. Both models are trained using knowledge distillation from the VideoPrism video foundation model. The models are evaluated in terms of recognition performance, parameter count, computational cost, inference latency, and throughput. The results show that knowledge distillation improves the recognition performance of both student models without changing their inference-time architecture. 3D-ResNet-18+BiLSTM achieves stronger recognition performance, while PoI-ViT requires fewer parameters and less computation and provides lower inference latency. The results also show that model selection should depend on the recognition and computational requirements of the target application.