Institutional Repository of Vilnius University
Automobilių atpažinimas iš vaizdo srauto /
Abstract
dc:descriptionVehicle recognition from video stream is a typical modern word problem, which can be used for calculating car speed, traffic light regulation and checking of insurance validity or technical inspection. This work describes vehicle dataset collection, preparation and ways of training image recognition model which gives the impression of needed tools, important facts, training parameters and possible improvements. Noted that it is neccessary to carefully test and remove all unnecessary information from dataset and since SSD algorithm has difficulty recognizing small objects, removing them also helps improve results. While training detection model many parameters were tried, including the use of different training hardware such as TPU and GPU which both have their own advantages and disadvantages. Use of TPU is faster and cheaper way to train your detection model which lets you try more different parameters and values. Best results on TPU were achieved using RMSProp optimization function, cosine learning rate, lower layers for feature extraction and the adjustment of scales that are determined using the KMeans clustering algorithm. To test detection model a program, which calculates passing vehicles by their category was written and a few possible difficulties and techniques present in this type of image analysis software have been noted. One of them is modern cameras and their fairly high resolution which could be reduced to increase inference speed.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kozakas, Rolandas,
- Contributors dc:contributor
-
- Grevys, Andrius
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD81918860&prefLang=en_US