Analýza semaforov v doprave pomocou obrazov

Abstract

The purpose of this thesis is to create an application, which can detect and analyze traffic lights in traffic through images. Motivation for this thesis is to get a closer look how autonomous vehicles detect objects in images. In images themself, we will detect the location of traffic lights and extract their state. We will use a method, which is currently considered as one of the best methods to detect objects in images, called detection with convolutional neural networks. In the first section of the thesis, we will describe the basic features of convolutional neural networks, layers of said network and dive in a bit of how this method of detection works. After that we will describe the dataset which we used and the process of collecting. Lastly we will describe the steps of training our own detector, compare tested architectures and also test the detector trained by us on a real word data.