Person re-identification of people and their luggage
Author
Nielsen, Simon Gørtz Flou
Term
4. term
Publication year
2022
Submitted on
2022-06-02
Pages
60
Abstract
Efterhånden som luftrejser bliver mere udbredte, har lufthavne brug for værktøjer til at overvåge passagerstrømme og anslå ventetider i køer. Person-genidentifikation (ReID) er en computer vision-teknik, der vurderer, om to kamerabilleder viser den samme person, og kan dermed støtte overvågning og kømålinger. Dette speciale undersøger person-ReID i denne sammenhæng ved at bruge billeder, hvor passager og medbragt bagage er markeret hver for sig. Datasættet indeholder to afgrænsningsbokse (rektangler): én omkring personen og én omkring bagagen. Med dette to-boks datasæt afprøver vi flere måder at udnytte informationen på. En Dual-TransReID-model, som behandler oplysninger om person og bagage i parallelle spor, opnåede 95% nøjagtighed i at matche identiteter. Det er en forbedring i forhold til en grundmodel (TransReID) med én samlet boks omkring både person og bagage, som opnåede 90% nøjagtighed. Koden findes her: https://github.com/MrFlou/Multi-Net-ReID
As air travel grows, airports need tools to monitor passenger flow and estimate queue times. Person re-identification (ReID) is a computer vision technique that assesses whether images from different cameras show the same person, supporting surveillance and queue measurements. This thesis investigates person ReID in this setting by using images where the passenger and the luggage they carry are marked separately. The dataset uses two bounding boxes (rectangles): one around the person and one around the luggage. Using this two-box dataset, we test several ways to leverage the information. A Dual-TransReID model, which processes person and luggage in parallel streams, achieved 95% accuracy in matching identities. This improves on a baseline TransReID model with a single box around both person and luggage, which achieved 90% accuracy. The code is available at: https://github.com/MrFlou/Multi-Net-ReID
[This summary has been rewritten with the help of AI based on the project's original abstract]
Documents