Thesis Details
Counting Vehicles in Image and Video
Traffic analysis is still a challenging task. During such task there are many pitfalls to be aware of. Such as small image resolution, high number of overlapping objects, angle of camera, blurred objects due to their motion or weather conditions. This thesis addresses these issues by using the convolutional neural network approach. In this thesis I propose a new architecture which adheres to Counting by Regression principle. The proposed architecture is inspired by some state-of-the-art architectures and improves accuracy on various datasets. For instance on the very small PUCPR+ dataset the Root Mean Square Error between expected and predicted vehicle counts was reduced from 34.46 to 6.99 vehicles (measured on the test set). Results achieved showed that there is still space for improvements and a possible further research in Counting by Regression principle.
vehicle counting, counting by regression, convolutional neural networks
Bidlo Michal, doc. Ing., Ph.D. (DCSY FIT BUT), člen
Čadík Martin, doc. Ing., Ph.D. (DCGM FIT BUT), člen
Grégr Matěj, Ing., Ph.D. (DIFS FIT BUT), člen
Kočí Radek, Ing., Ph.D. (DITS FIT BUT), člen
@bachelorsthesis{FITBT23010,
author = "Dominik Gabzdyl",
type = "Bachelor's thesis",
title = "Counting Vehicles in Image and Video",
school = "Brno University of Technology, Faculty of Information Technology",
year = 2020,
location = "Brno, CZ",
language = "english",
url = "https://www.fit.vut.cz/study/thesis/23010/"
}