Anomaly Detection in Logistics Warehouses Based on YOLOv8

Due to the rapid growth in demand for the logistics industry, efficiently handling anomaly events in logistics warehouses and thereby enhancing work efficiency and security has become an important issue. In this paper, four types of anomaly events have been defined, each corresponding to a firsthand industry dataset collected from a local logistics warehouse. By comparing the definitions of the anomaly events with the behaviors observed in the videos, an alarm will be sent if they match, warning the workers to take measures to avoid further negative impacts. To be more specific, four models were trained using the specific datasets and the You Only Look Once Version 8 (YOLOv8) network—the latest version of the YOLO series known for its strong performance and lightweight structure—to perform object detection and instance segmentation tasks. Furthermore, anomaly analysis algorithms were integrated into the four models to analyze deviations of objects and instances observed in the videos from the norm. In this case, the anomaly detection algorithms employed a two-stage process to automatically identify anomalies. The experimental results and deployment simulation outcomes indicate that, in a GPU environment, these algorithms achieve competitive performance in terms of accuracy and speed, making them suitable for integration into a real-time intelligent video surveillance system.