To address the decline in pedestrian detection accuracy caused by complex scenarios such as illumination variations, viewing angles, background interference and small pedestrian targets, which often lead to false positives and missed detections, a pedestrian detection model, YOLOv11-CREP, was proposed based on an improved YOLOv11n. Firstly, CSPDConv, which was formed by integrating standard convolution(Conv) with space-to-depth convolution(SPDConv), was introduced to reduce information loss and enhance critical feature extraction. Secondly, a new RepNCSPELAN4-GC module was proposed, which incorporates GhostConv to optimize the RepNCSPELAN4 module, reducing its parameter count. The improved RepNCSPELAN4-GC module was then used to partially replace the C3k2 modules in the Neck layer. Next, efficient multi-scale attention(EMAttention) and parallel network attention(ParNetAttention) were fused into a new EMPAttention module to enhance the detection ability of the model for small target pedestrians. Finally, considering the characteristics of small target pedestrains and occluded targets, a small-target detection head P2 was added to further improve the model’s recognition capability for small targets. The experiments show that compared with the original YOLOv11n model, YOLOv11-CREP improves the mean average precision(mAP) by 4.6 percentage points at an IoU threshold of 0.5, reaching 95.3%. When evaluated over the IoU range of 0.5 to 0.95, its mAP increases by 9.0 percentage points, reaching 70.2%. The proposed model achieves a balance between high detection performance and real-time requirements, effectively enhancing pedestrian detection performance in complex scenarios. It provides valuable references for modeling pedestrian detection tasks.