【摘要】 在公共场合，异常斗殴行为对公共安全来说是极大的隐患。为了配合监控实时检测人的异常行为，提出一种基于YOLOv8n的实时精准公共场合异常行为检测算法——Swin-YOLO。首先，在主干网络将C2f模块替换为能够提高特征融合度的Swin-Conv模块，以改善主干网络的特征融合效率；其次，在主干网络中插入ParNet注意力机制，以更好地提取丰富的全局信息。由于缺少专门针对公众场合斗殴行为的公开数据集，本文通过自建数据集与公共数据集结合进行实验。实验结果表明：相比原有的基础YOLOv8n目标检测模型，改进的算法的mAP50综合指标达到了87.8%,提升了2.5%。同时，改进的算法能够保持1 ms的推理速度，在实际应用中达到了94.8%的检测率，能够符合实时检测的要求。更多还原

【Abstract】 Unusual fighting behavior in public places is a great risk to public safety.To cooperate with surveillance to detect human’s abnormal behavior in real-time, Swin-YOLO,a real-time and accurate algorithm for detecting abnormal behaviors in public places based on YOLOv8n is proposed.Firstly, the C2f module in the backbone network is replaced with the Swin-Conv module, which is capable of improving feature fusion to improve the feature fusion efficiency of the backbone network.Secondly, ParNet attention mechanism is inserted into the backbone network to extract the rich global information better.Due to the lack of a public dataset dedicated to public fighting behavior, this paper conducts experiments by combining a self-constructed dataset with public datasets.Experimental results show that the mAP50 composite index of the improved algorithm reaches 87.8 %,which improves 2.5 %,compared with the original basic YOLOv8n target detection model.At the same time, the improved algorithm is able to maintain the inference speed of 1 ms, and achieves a detection success rate of 94.8 % in practical applications, which can meet the requirements of real-time detection.更多还原