Railway Foreign Object Intrusion Detection Using UAV Images and YOLO-UAT

Abstract

Rapidly detecting foreign objects in the railway’s natural environment is crucial for safe railway operation and passenger safety. Traditional detection methods are limited by technology and environment and are costly and inefficient for large-scale detection. To improve the efficiency of foreign object intrusion detection in railways, this study proposes a YOLO-UAT method to detect foreign objects in railways effectively. First, a railway foreign object intrusion image dataset is created. Then the EfficientNet network is employed to replace the backbone extraction network of YOLOv5s, achieving a lightweight model that enhances detection speed. Secondly, a C $3\_$ CBAM module is constructed to enhance feature extraction and enhance the model’s detection ability for small-scale targets. Concurrently, the K-means++ algorithm is introduced to cluster the a priori frames, which improves the accuracy and convergence speed of the a priori frame clustering. YOLO-UAT reduces the number of parameters by 36% compared to the original YOLOv5s and mAP increased by 6.1% to 91.5%. Implemented on a Jetson Nano, it achieves a detection rate of 26.4 FPS. The experimental results demonstrate that the enhanced YOLOv5s model effectively enhances the accuracy and speed of foreign object encroachment detection in railroads while ensuring lightweightness, facilitating deployment, and aligning with the operational needs of railroads.

Keywords

• Object detection
• UAV
• YOLOv5s
• EfficientNet
• lightweight
• CBAM