A Foreign Object Detection Method for Railway Overhead Lines Based on Few-Shot Learning

It often occurs that foreign objects hang on railway overhead lines especially in windy days, which may lead to the delay of train, and even cause loss of life and property. With the gradual maturity of computer vision, railway foreign object detection has made great progress. This paper focuses on foreign object in railway overhead lines, and proposes a foreign object detection method based on few-shot learning. The proposed method consists of two stages: general training and fine-tuning training. In the first stage, the data-abundant common classes dataset is used to train the entire detector to obtain the underlying model parameters. In the fine-tuning stage, we make a balanced dataset consisting of common classes and enhanced railway classes, and fine-tune the last layer of the detector to accomplish few-shot foreign object detection. Finally, this proposed method is tested on images of overhead lines intrusions taken in actual railway scenario, and achieves 79% mAP in railway classes.