A nighttime pedestrian detection method based on YOLOv8s with joint image enhancement and network structural optimization

Accurate pedestrian detection in nighttime traffic scenarios remains a critical challenge for autonomous driving perception systems, mainly due to contrast degradation caused by insufficient illumination and the loss of target information caused by occlusion. To address these issues, this study adopts YOLOv8s as the baseline detector and proposes a joint framework that integrates image enhancement and structural optimization for nighttime pedestrian detection. Specifically, an I-SCI illumination enhancement strategy is embedded into the front-end of the improved YOLOv8s detector to enhance the visibility of low-light images. Furthermore, three structural optimization strategies are introduced to overcome the challenges of poor illumination and target occlusion. First, RepGFPN is employed to construct multiple cross-layer feature interaction paths, enhancing multi-scale feature fusion capability and enabling the network to better distinguish pedestrian contours from background light interference. Second, a Dynamic Attention Transformer (DAT) module is incorporated to focus on critical spatial response regions. By leveraging contextual information from visible body parts, such as the head and legs, DAT can infer the semantic information of occluded regions, thereby reducing missed detections and suppressing interference from complex backgrounds. Third, the MPDIoU loss function is introduced to improve localization accuracy. Through refined corner constraints, MPDIoU effectively reduces prediction box deviation and enhances the spatial consistency between predicted and ground-truth bounding boxes. The proposed optimization strategies work collaboratively to improve the accuracy and robustness of pedestrian detection under low-light conditions. Experimental results demonstrate that YOLO-PN achieves significant improvements over the baseline YOLOv8s, with Precision, Recall, and mAP increased by 4.15, 4.57, and 3.88% points, respectively. These results indicate that the proposed method substantially enhances the adaptability and detection stability of the YOLOv8s framework in complex nighttime traffic scenarios, providing an effective solution for nighttime perception in autonomous driving systems.