A Deep Learning-Based Flight Turnaround Record System

To ensure the smooth progression of the flight turnaround, airports need to record the start and end times of each node in the flight turnaround process, compare them with the desired schedule, and guide the progress of subsequent flight tasks. Therefore, the timeliness and accuracy of data collection at each turnaround node become crucial. Traditionally, some airports manually record flight turnaround node times, which is inefficient and unable to realize data structuring and visualization. To address the above issues, we propose a flight turnaround record system based on object detection. The system builds upon the YOLOv5s-based detection network by introducing attention mechanisms with FasterNet as the backbone to enhance model performance and inference speed. Furthermore, we introduce dual-perspective coordination for aircraft positions, adaptive masking, and other methods to assist the system in making logical determinations about the status of nodes. Experiments have shown that our proposed detection network outperforms the baseline YOLOv5s method on our self-built apron dataset, with a 20fps inference speed increase and a 0.08 mAP@.5:.95 detection accuracy improvement. Compared to manual methods, our system reduces the average time error for each node by 60–120 seconds, significantly improving the efficiency of flight turnaround.