Is larger always better? A comprehensive evaluation of deep learning models for foreign object detection in metro systems

Efficient and automatic foreign object detection (FOD) between platform screen doors (PSDs) and metro doors is crucial for intelligent metro operations. While deep learning has demonstrated exceptional performance in object detection tasks, the deployment of large models in metro systems presents significant practical challenges due to their computational demands. This study investigates the optimal balance between detection performance and operational feasibility in metro FOD applications. A systematic analysis of FOD challenges in metro environments is first conducted, identifying key issues including detection uncertainty, data constraints, and computational limitations. Through collaboration with Guangzhou Metro Group Co., Ltd., (Guangzhou, China) the first large-scale metro FOD dataset was established, comprising 5,854 images with diverse foreign objects from real operational scenarios. Then, 36 different object detection algorithms were evaluated, ranging from large-scale models to lightweight architectures, focusing on their practical deployment capabilities. The comprehensive experiments reveal that lightweight neural networks, particularly YOLOv5-s, achieve superior practical performance in metro environments. While larger models demonstrate marginally higher detection accuracy (YOLOv5-x: 0.894 mAP), lightweight alternatives offer substantially better deployment value through balanced accuracy (YOLOv5-s: 0.880 mAP), real-time processing capability (588 FPS), and efficient resource utilization (13.7 MB). These findings provide valuable guidance for implementing deep learning-based FOD systems in real-world metro operations.