Pedestrian detection in low light conditions

Thumbnail Image

Files

master_Bouzoulas_Dimitrios_2025.pdf (20.95 MB)

URL

Journal Title

Journal ISSN

Volume Title

School of Engineering | 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

2025-07-30

Department

Major/Subject

Mechanical Engineering

Mcode

Degree programme

Master's programme in Mechanical Engineering

Language

en

Pages

52

Series

Abstract

Pedestrian detection in RGB images is a key task in pedestrian safety, as the most common sensor in autonomous vehicles (AVs) and advanced driver assistance systems (ADAS) is the RGB camera. RGB pedestrian detection is challenging in low light conditions, and at the same time, the probability of a fatal accident is high according to statistics. To enable RGB pedestrian detection at night, large labelled datasets are required, yet most publicly available datasets consist primarily of daytime conditions. This thesis aimed to solve this problem by providing a pipeline capable of automatically labelling nighttime RGB data. For this goal, RGB-Infrared image pairs are used. This solution leverages the lighting invariance of infrared cameras to 1) Detect pedestrians on the infrared portion of the dataset, this step if followed by 2) Label transfer from infrared to RGB and finally, 3) Model training with the annotated low light RGB dataset. The research was performed using the KAIST dataset. For the automatic labelling, fine-tuned models capable of detecting pedestrians in infrared images were used. For the evaluation, object detection models were trained on the generated autolabels and ground truth labels. When compared on a previously unseen image sequence, the results showed that the models trained on generated labels outperformed the ones trained on ground-truth labels in 6 out of 9 cases for the mAP@50 and mAP@50-95 metrics. The pipeline presented some positive initial results. Next, it has to get further evaluated by performing experiments on a larger scale and with different datasets, as it was only evaluated with KAIST. The successful further evaluation would enable large-scale automated labelling, significantly aiding the training on low light RGB object detection models. Finally, a separate task is the exploration of better pedestrian detection models for infrared pedestrian detection, as the model has the abilty to detect pedestrians in infrared images is directly correlated with label quality. The source code for this research is available at: https://github.com/Bouzoulas Dimitrios/IR-RGB-Automated-LowLight-Pedestrian-Labeling

Description

Supervisor

Ojala, Risto

Thesis advisor

Ojala, Risto

Keywords

computer vision, autonomous vehicles, pedestrian detection, low light detection, pedestrian safety, automatic labelling

Other note

Citation

Permanent link to this item

https://urn.fi/URN:NBN:fi:aalto-202508196323

Collections

[dipl] Insinööritieteiden korkeakoulu / ENG