Enhancing low-light pedestrian detection: convolutional neural network and YOLOv8 integration with automated dataset

Abstract

This research aims to enhance the you only look once (YOLO) model for pedestrian detection in environments with varying lighting conditions, particularly in low-light scenarios. The primary contribution of this work is the integration of a convolutional neural network (CNN)-based low-light enhancement model, which transforms dark images into brighter, more discernible ones. This enhanced dataset is subsequently used to train the YOLO model, allowing it to learn from both the original and transformed data distributions. Unlike traditional YOLO training approaches, this method generates more accurate data representations in challenging lighting environments, leading to improved detection outcomes. The novelty of this approach lies in its dual-stage training process, which integrates a CNNbased low-light enhancement model with YOLO’s detection capabilities. This combination not only enhances pedestrian detection but also has the potential for application in other domains, such as vehicle detection and surveillance, particularly in challenging lighting conditions. The automatic dataset collection pipeline provides an efficient way to gather diverse training data across various scenarios. The YOLOv8 model trained on the low-light enhanced dataset significantly outperformed the baseline model trained only on the original dataset, with precision increased by 9.8%, recall by 45.7%, mAP50 by 26.8%, and mAP50-95 by 41.0% when validated on dark images.