SwinIR-YOLOv8 Pedestrian Detection In Adverse Weather Conditions

Aiming at the problem of poor image quality and severe noise interference under adverse weather conditions, which leads to low pedestrian detection accuracy, this paper proposes a pedestrian detection algorithm SwinIR-YOLOv8 that can be used in adverse weather conditions to ensure that each input image can be adaptively enhanced to achieve better detection performance. Specifically, the SwinIR image enhancement module is introduced for image restoration, and to better preserve more feature information, the GAM attention mechanism is introduced to dynamically combine feature networks from different receptive fields, enhancing the algorithm's processing ability in harsh environments. We conducted comparative experiments on improved algorithms such as IA-YOLO and AOD-YOLOv5 for severe weather conditions. The experimental results showed that the proposed algorithms improved accuracy by 1.61% and 2.24%, respectively, verified that this algorithm can effectively improve the accuracy of pedestrian detection in harsh environments.