What is it about?
To create intelligent traffic-light control in cities with many pedestrians and cyclist, fine-grained tracking of those road users is needed. To train those algorithms labeled data-sets are needed, which unfortunatelly mainly available during daylight. To create additional training-data at night, the daylight recordings can be augmented to the lighting conditions at night. This study shows that when done in the right way, the robustness of the algorithms could improve substantial.
Featured Image
Photo by Dima Pechurin on Unsplash
Why is it important?
A city can improve its welfare when pedestrians and cyclist are supported with intelligent traffic lights. This should not only work under good lighting conditions, but also at night.
Perspectives
I like two things from this study. Firstly, the quality of the augmentation (from day to night conditions) is measured with the failure rates of the tracking algorithm. So, the night-images not only have to look realistic, but also need to contribute as good training material. Secondly, the quality boost for the tracking algorithm is not so much in a few percentage-points higher accuracy or precision, but was mainly visible in the robustness. The algorithm was able to provide the same tracking results for a larger set of scenarios.
Dr. Arnoud Visser
Universiteit van Amsterdam
Read the Original
This page is a summary of: Improving Multi-Object Re-identification at Night with GAN Data Augmentation, January 2024, Springer Science + Business Media,
DOI: 10.1007/978-3-031-44851-5_37.
You can read the full text:
Contributors
The following have contributed to this page
