Smart Traffic Control Using Optimised Convolutional Neural Network

dc.contributor.advisorAnbarjafari, Gholamreza
dc.contributor.authorSurrage Reis, Mateus
dc.date.accessioned2019-06-04T12:27:14Z
dc.date.available2019-06-04T12:27:14Z
dc.date.issued2019
dc.description.abstractIn English: The state-of-the-art in image object detection is in convolutional neural networks, which is a computationally expensive base to build on. To run accurate detection in an embedded device, additional optimization is required if there is a need to run it real-time on each frame of a video. This thesis details work done in the development of a smart pedestrian crosswalk: an Internet of Things enabled embedded platform for traffic control. By fine-tuning an individual neural network for each SPC post, it was possible to significantly boost accuracy in a fast, low-accuracy CNN. This was accomplished by taking advantage of the low variation in possible input images, being drawn from only 3 cameras per post. The improvement was from 33.1% mAP in general context images and 80 classes to 60.7% mAP on solely traffic images and seven traffic-relevant classes. Eesti keeles: Uusimad objekti tuvastus meetodid kasutavad oma töös konvuleerivaid närvivõrke, mis arvutuslikust küljest on ressursiahned. Täpsete tuvastusmudelite jooksutamine manussüsteemides vajab palju optimeerimist, eriti kui seade peab toimima reaalajas. Käesolev töö kirjeldab targa ülekäiguraja teemärgi loomist: nutiseade, mis on mõeldud liikluse juhtimiseks. Seadistades iga SPC posti eraldi närvivõrku, oli võimalik märkimisväärselt tõsta kiire ja ebatäpse närvivõrgu selgust. See saavutati kasutades ära kolmest kaamerast tulevate sisendpiltide minimaalset varieeruvust. Algoritmi täpsust parandati 80 klassilise üldnärvivõrgu 33.1% mAP pealt 60.7% mAP peale, rakendades ainult liiklusega seotud pilte koos seitsme erineva teemakohase klassiga.en
dc.identifier.urihttp://hdl.handle.net/10062/63951
dc.language.isoenget
dc.publisherTartu Ülikoolet
dc.rightsembargoedAccesset
dc.subjectCNNen
dc.subjecttehisnärvivõrket
dc.subjectnutistuet
dc.subjectautonoomneet
dc.subjecttehisintellektet
dc.subjectneural networksen
dc.subjectIoTen
dc.subjectautonomousen
dc.subjectAIen
dc.titleSmart Traffic Control Using Optimised Convolutional Neural Networken
dc.title.alternativeTark liikluse juhtimine rakendades optimeeritud konvuleerivat närvivõrkuet
dc.typeThesiset

Files

Original bundle

Now showing 1 - 1 of 1
Loading...
Thumbnail Image
Name:
Surrage Reis_BSc2019.pdf
Size:
3.51 MB
Format:
Adobe Portable Document Format
Description:

License bundle

Now showing 1 - 1 of 1
No Thumbnail Available
Name:
license.txt
Size:
1.71 KB
Format:
Item-specific license agreed upon to submission
Description: