在密集场景中，常常包含众多被遮挡或者小尺度的行人目标。这样的场景对常规的目标检测模型提出了挑战，往往会出现大量的漏检和错检问题。为了解决密集场景中行人检测时出现的高漏报率和误报率问题，提出了一种新的密集场景行人检测框架ST-YOLO。首先，将YOLOv5的骨干网络中的低层小目标检测层融入特征金字塔网络和路径聚合网络结构中，增加了一个检测小目标行人检测层；其次，对YOLOv5的颈部网络进行改进，利用基于Swin Transformer的多尺度全局信息和卷积神经网络（convolutional neural networks, CNN）所提取的局部信息来构建聚合特征，提高网络的特征提取能力；并且在预测过程中引入了SIoU（scalable-IoU）损失函数，加快模型的收敛速度和提升检测能力；最后，使用Soft-NMS（soft non-maximum suppression）代替原非极大值抑制（non-maximum suppression, NMS）算法，减少非最大化抑制阶段误删除检测框问题，降低了检测算法的误报率。在Wider Person数据集上的大量实验表明，改进后的ST-YOLO算法的精度和mAP_0.5比目前主流的YOLOv9算法分别提升了5.7%和3.6%。

In dense scenes, the frequent occurrence of occluded or small-scale pedestrian objects poses significant challenges to traditional object detection models, frequently leading to a high number of missed detections and false positives. In order to solve the problem of high false negative rate and false positive rate in pedestrian detection in such dense scenes, a novel dense scene pedestrian detection framework called ST-YOLO was proposed. Firstly, the low-level small object detection layer in YOLOv5's backbone network was integrated into the feature pyramid network and path aggregation network structure, adding a pedestrian detection layer for detecting small objects. Secondly, the neck network of YOLOv5 was improved by utilizing multi-scale global information based on Swin Transformer and local information extracted by convolutional neural networks (CNN) to construct aggregated features and enhance the network's feature extraction capability. And the SIoU (scalable IoU) loss function was introduced in the prediction process to accelerate the convergence speed of the model and improve detection capability. Finally, Soft NMS (soft non maximum suppression) was used instead of the original non maximum suppression (NMS) algorithm to reduce the problem of mistakenly deleting detection boxes during the non maximum suppression stage and lower the false alarm rate of the detection algorithm. A large number of experiments on the Wide Person dataset have shown that the improved ST-YOLO algorithm has improved accuracy and mAP_0.5 by 5.7% and 3.6% respectively compared to the current mainstream YOLOv9 algorithm.