Adaptive YOLOv6-Driven Infrared Pedestrian Perception for Low-Visibility Intelligent Transportation Systems

Abstract

Reliable pedestrian detection during nighttime is a fundamental requirement for intelligent transportation systems and autonomous driving because poor illumination often reduces the effectiveness of conventional vision-based detection methods. This paper presents a deep learning-based framework for nighttime pedestrian detection by evaluating the performance of Faster R-CNN, an enhanced YOLOv5 model, and an extended YOLOv6 architecture using infrared pedestrian images. The experimental study utilizes the CVC-09 infrared pedestrian dataset containing 2,199 images. The dataset is preprocessed through image shuffling, annotation handling, and partitioning into training and testing sets before model development. The enhanced YOLOv5 model incorporates an additional attention-based squeeze mechanism to improve feature extraction, while the extended YOLOv6 architecture is employed to further strengthen detection capability. The developed models are assessed using accuracy, precision, recall, and F1-score, together with graphical analysis of training accuracy and loss. Experimental findings indicate that the baseline Faster R-CNN achieves an accuracy of approximately 77%, whereas the enhanced YOLOv5 model improves the detection accuracy to about 96%. The extended YOLOv6 model delivers the best performance with nearly 99% detection accuracy, demonstrating superior localization and classification of pedestrians in infrared nighttime images. The proposed framework provides an efficient and robust solution for nighttime pedestrian detection and can be integrated into advanced driver assistance systems and autonomous vehicles to improve road safety under low-visibility conditions