Analýza okolních automobilů pro autonomní vozidlo za pomocí obrazů

Abstract

The objective of this bachelor thesis is to investigate and evaluate the functionality and accuracy of selected vehicle detection solutions. For comparison purposes, several models were trained on a modified dataset from Udacity and on my own. The algorithms used were YOLOv5, ResNET and VGG, implemented in PyTorch. The first part deals mostly with the basics and theory. It describes possible use cases for vehicle detection, what detection methods are suitable for that particular use case, and examples of use cases. Towards the end of the theoretical part, it describes what a dataset is, a description of the datasets used for the practical part and their comparison. The second (practical) part deals with the implementation of vehicle detection models using the publicly available YOLOv5 library and the subsequent comparison of detection success rates for different sized networks and using different datasets. Finally, I mention the possibilities of extending the methods with a classification along with a comparison of different approaches.