基于改进YOLOv8的监控中异常行为检测算法研究  

Research on Abnormal Behavior Detection Algorithm in Monitoring Based on Improved YOLOv8

作　　者：柴威 杨明浩 韩沈飞 何芳州[1] CHAI Wei;YANG Minghao;HAN Shenfei;HE Fangzhou(Public Security Information Technology and Intelligence College,Criminal Investigation Police University of China,Shenyang 110854,China)

机构地区：[1]中国刑事警察学院公安信息技术与情报学院,沈阳110854

出　　处：《刑事技术》2025年第2期124-131,共8页Forensic Science and Technology

基　　金：辽宁省社会科学规划基金项目(L21ASH004);智能警务四川省重点实验室项目(ZNJW2022KFMS002)。

摘　　要：异常行为检测是维护治安的一项重要任务。快速准确地对关键场所中的异常行为进行检测是治安工作的重要任务。对于关键场所中的目标检测算法来说,要在人员密集的图像中获得满意的检测结果具有一定挑战性。本文针对异常行为目标检测中存在的目标分布密集、尺度变化范围较大、检测背景较为复杂等问题,将注意力机制CA引入到主干网络部分C2f模块中,从而加强网络在复杂背景下对目标的注意力;将Swin Transformer模型集成到YOLOv8骨干网络中,增加特征图的信息交互范围,充分利用物体的背景信息提高复杂背景下物体的检测精度。最后通过实验验证了改进网络的可行性,并与几种主流网络进行对比,平均精度达到95.1%,与基础网络YOLOv8相比,平均精度提高2.4个百分点,证明了该方法的有效性。Detecting abnormal behavior is crucial for maintaining public security,especially in densely populated critical areas.Traditional target detection algorithms often struggle to deliver satisfactory results under these conditions due to challenges like dense target distribution,signifi cant scale variation,and complex backgrounds.YOLOv8 is one of the better perforing detection models effect among the object detection models.This study introduces a novel approach to improve detection accuracy by integrating advanced mechanisms into the YOLOv8 backbone network.Firstly,the coordinate attention(CA)mechanism is incorporated into the C2f module of the backbone network.This enhances the network’s focus on targets amidst complex backgrounds by emphasizing relevant features and suppressing noise.Secondly,the swin transformer model is integrated into the YOLOv8 backbone.The swin transformer facilitates greater information interaction across the feature map,effectively utilizing the background information and improving object detection accuracy under complex scenarios.The datasets used in the experiments are described,the evaluation indexes of P,R,AP and mAP are listed,and ablation experiments and comparative experiments are carried out.Experiments demonstrate the feasibility and effectiveness of these improvements.The enhanced network is compared with several mainstream networks,showing a signifi cant improvement in average accuracy,reaching 95.1%.Compared to the basic network YOLOv8,the average precision has been improved by 2.4%,which proves the effectiveness of this method.In summary,the innovative integration of the CA mechanism and Swin Transformer model into the YOLOv8 backbone network addresses key challenges in detecting abnormal behavior in densely populated and complex environments.These enhancements lead to improved detection accuracy,making it a promising approach for public security applications.

关 键 词：异常行为 目标检测 YOLOv8 深度学习 注意力机制 

分 类 号：DF793.2[政治法律—诉讼法学]