Dense crowd object detection, as a means to monitor security risks and crowd control in densely populated large venues, is a challenging research area. However, issues such as mutual occlusion between objects or large differences in scale between foreground and background objects lead to problems like missed detections and false alarms in pedestrian recognition within dense crowds. To address these challenges, this paper proposes a dense crowd object detection method based on YOLOv8-MGH. We optimized the FPN feature pyramid module of the YOLOv8 algorithm into a Generalized Enhanced Feature Pyramid Network (GEFPN) to achieve feature fusion across different levels and retain human features despite occlusion. To address the issue of significant scale differences between foreground and background objects, we introduced a Multi-Scale Channel Split module (MCS) to enhance the model’s ability to perceive human features at various scales. Additionally, we added a Tiny Object Detection Head (TOD-Head) to the head, significantly improving the model’s ability to detect tiny objects and increasing detection accuracy. We conducted experiments using both large parameter and small parameter versions on the publicly available CrowdHuman dataset. The results demonstrate that our method outperforms state-of-the-art methods in both versions. Additionally, we validate the model’s generalization on the Widerperson dataset.