Perception of passenger abnormal behavior in metro stations based on improved YOLOv8
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.
万方数据
维普
