Road marking condition monitoring and classification using deep learning for city of Helsinki.

Thumbnail Image

Files

master_Tokmurzina_Dana_2020.pdf (12.04 MB)

URL

Journal Title

Journal ISSN

Volume Title

Perustieteiden korkeakoulu | Master's thesis
Unless otherwise stated, all rights belong to the author. You may download, display and print this publication for Your own personal use. Commercial use is prohibited.

Authors

Date

2020-10-20

Department

Major/Subject

Data Science

Mcode

SCI3095

Degree programme

Master's Programme in ICT Innovation

Language

en

Pages

60+6

Series

Abstract

The thesis explores application of deep learning on detection and classification of road markings in the city of Helsinki. The need for maintaining the infrastructure is the essential part for smart cities. City of Helsinki is thriving towards the digitization of the city, providing geo spatial information on one of their open geoinformatics service, https://kartta.hel.fi. That was utilized as the main data source. Based on the satellite images obtained from kartta.hel, the road markings were extracted. Further, previous work on zebra crossings was studied, both using traditional ways and deep learning (DL) based ones. Deep Learning were favoured over traditional due to ability to capture deeper abstract concepts and hierarchical features. Several recent DL based object detection algorithms, their training process, hyperparameter tuning, results are described in depth. In addition history of computer vision, especially object detectors, their benchmarks, disadvantages, and advantages are studied extensively. Taking into account the specifics of the dataset such as low resolution, small size and noise, data augmentation and transfer learning were applied. After the comparison between various object detection algorithms and also taking into account requirements for the performance as accuracy, robustness to noise, shadows, state of the art algorithms were chosen, such as Retina Net and YOLO5. YOLO5 outperformed in all desired metrics. It achieved mAP\_0.5 of 0.68, inference time of 0.017 seconds with relatively low (compared to RetineNet) time for training. In addition it produced good visual results on the test dataset.

Description

Supervisor

Jung, Alexander

Thesis advisor

Lohi, Saska

Keywords

machine learning, deep learning, object detection, image processing, RetinaNet, YOLO5

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202011016271

Collections

[dipl] Perustieteiden korkeakoulu / SCI