Institutional Repository of Vilnius University
Information system with automatic vehicle recognition /
Abstract
dc:descriptionNowadays, the number of four-wheeled vehicles on the road has increased dramatically over the past couple of decades. Automobiles have become more affordable for mid to low-income families, and some even have a vehicle per person. This leads to an explosive number of varieties that can be seen daily, and with new vehicles on the way, it becomes difficult to identify some models or differentiate them from one another. Although with the quite recent uprise of Artificial Intelligence, it is now easier to separate the differences of appearance, there are still some drawbacks of not enough mainstream models that deliberately focus on the categorization of these vehicles for various purposes, largely known as camera recognition, excluding number plates. Therefore, there is a need to develop a reliable sole-purpose computerized model that automatically recognizes different vehicles for convenience and enhanced automation. The goal of this work was to research and investigate the performance of various existing CNN models on four-wheeled vehicles and develop an enlarged custom dataset, which then can be used to train different fine-grained models, compare, and draw conclusions about which one is the most effective one for use in the real world. After this, the most advanced models were implemented into a created prototype web application with the functionality of testing the models via web interface and API keys. The Stanford Cars dataset, together with the ResNet18 ML model, was used first, laying the foundation for the vehicle recognition system. Afterward, different types of augmentation were implemented to see the impact on accuracy in the trained models. Later on, ResNet50 was exchanged instead of ResNet18 to see the effects of more complex and in-depth features, such as many more parameters. Finally, the VMMR dataset was connected with the Stanford Cars to create a unified and bigger dataset. The Stanford Cars dataset and the unified dataset were separately evaluated on the EfficientNet-b4 ML model. This specific model was chosen due to its even deeper architecture and better ability to model fine-grained visual details. The training and validation results were then compared to locate the most efficient model. The best performing models, which were concluded to be both trained with the EfficientNet-b4 architecture, were implemented into the later-created prototype web application.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Skačkauskas, Kasparas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD277690318&prefLang=en_US