基于改进YOLOv8的地铁站内乘客异常行为感知

Perception of passenger abnormal behavior in metro stations based on improved YOLOv8

当地铁站内乘客出现异常行为时,若未能及时发现可能会引起乘客不满、投诉,甚至导致安全问题,从而影响运营效率,造成恶劣影响.而当前常用的盯控视频画面的方式存在容易遗漏和效率低的问题.为及时感知异常行为,提出一种云边协同的异常行为感知总体架构.首先,通过人工演绎的方法在地铁站内采集异常行为图像,构造包含11种异常行为的数据集;其次,针对边/端侧能够自主训练和推理但算力较小的特点,提出模型压缩算法,构建MINI-BLOCK模块并将其组合为i-C2f模块,用于替换YOLOv8中的C2f模块;再次,针对云侧计算资源集中的特点,分别构建2个基于YOLOv8的改进模型,即ModelA和ModelB,ModelA的架构为“DCNv2_Dynamic-BiFPN-EMA”,ModelB的架构为“DCNv2-BiFPN-EMA”;最后,在构造的数据集上,对提出的3种优化模型与YOLOv8进行对比实验.研究结果表明:相较于YOLOv8,3种优化模型均取得了更优的检测性能,边/端侧模型的精确率提升了1.0%,模型参数降低了4.7%;ModelA的召回率、mAP50、mAP50:95分别提升了2.2%、3.7%、2.9%;ModelB的召回率、mAP50、mAP50:95分别提升了5.8%、6.7%、2.8%.研究结果能够为地铁乘客异常行为感知的相关研究提供参考.

Timely detection of passenger anomalies in metro stations is crucial to prevent dissatisfac-tion,complaints,and potential safety hazards,thereby impacting operational efficiency and public safety.Current surveillance methods,reliant on constant video monitoring,often suffer from oversight and inefficiency.To address this,a novel cloud-edge collaborative architecture is proposed for abnor-mal behavior perception.Initially,images of abnormal behaviors in metro stations are collected using artificial enactment,forming a dataset with 11 anomaly types.To accommodate the limited computational power of edge devices,a model compression algorithm is developed,featuring the MINI-BLOCK module integrated into an i-C2f module,replacing the C2f module in YOLOv8.Fur-thermore,leveraging centralized cloud computational resources,two improved models are developed based on YOLOv8:ModelA with DCNv2_Dynamic-BiFPN-EMA architecture,and ModelB with DCNv2-BiFPN-EMA architecture.Finally,comparative experiments are conducted on the con-structed dataset among three optimized models and the original YOLOv8.The findings indicate that all three optimized models outperform YOLOv8 in detection capabilities.The edge-side model achieves a 1.0%increase in precision and a 4.7%reduction in model parameters.ModelA demon-strates a 2.2%improvement in recall,a 3.7%increase in mAP50,and a 2.9%enhancement in mAP50:95,while ModelB sees a 5.8%increase in recall,a 6.7%improvement in mAP50,and a 2.8%increase in mAP50:95.These results provide valuable insights for future research in metro pas-senger anomaly perception.