Attention-Enhanced Feature Fusion for Robust Pedestrian Detection in Dense Crowded Scenes

Pedestrian detection in dense crowded scenes remains a challenging problem due to severe occlusion, scale variation, and complex background interference. To address these issues, this paper presents an enhanced YOLOv8-based framework that improves feature representation and multiscale information aggregation for robust pedestrian detection. Specifically, an attention-guided convolution mechanism is introduced to enlarge the effective receptive field and strengthen contextual awareness, while an adaptive multi-level feature fusion strategy is employed to better integrate spatial and semantic information across different scales. In addition, an improved localization and suppression scheme is adopted to alleviate bounding box redundancy and enhance detection stability in highly overlapped scenarios.