Dirt Video Anomaly Detection Method Based on Automatic Identification of the Exit Port

This study proposes an online approach for detecting abnormalities in slag discharged from the shield machine using video anomaly detection and automatic identification of the soil outlet. By analyzing various soil anomalies, the method characterizes these irregularities. An automatic identification system based on You Only Look Once v3 is developed to process surveillance video while mitigating environmental interference. An adaptive frame sampling method is introduced to accurately identify normal and abnormal slag output speeds. Additionally, the 3DResnet50 network model is modified to effectively capture slag motion information. Experimental results show that this approach increases video classification accuracy to 98.81%, meeting online detection requirements for short video segments.