LE-G2F-YOLOv8: An Infrared Pedestrian Detector with Low-Frequency Enhancement and lightweight

Abstract

Pedestrian detection is a key task in computer vision with wide applications in intelligent surveillance and autonomous driving. Traditional visible-light-based methods often fail under low-light conditions due to poor contrast and occlusion. To address these challenges, this method proposes LE-G2F-YOLOv8, an infrared pedestrian detection model that integrates a Low-Frequency Enhancement (LEF) module and a lightweight G2F residual structure. The LEF module enhances the edge and semantic features of infrared images by amplifying low-frequency components, while the G2F block reduces redundant computation and accelerates inference. The proposed model is evaluated on the FLIR dataset, achieving significant improvements over the baseline YOLOv8 in terms of precision, recall, average precision (AP), and inference speed. Specifically, the model improves precision by 2.11%, recall by 3.39%, AP by 1.95%, and inference speed by 10.7%. These results demonstrate the effectiveness of LE-G2F-YOLOv8 for real-time pedestrian detection in low-illumination scenarios, showing strong potential for deployment in practical applications such as nighttime traffic monitoring and mobile robot perception systems.