A comparative study of machine vision-based rail foreign object intrusion detection models

Due to the lack of track foreign body intrusion dataset, classical target detection models are rarely used in the field of foreign body intrusion on railway tracks, and model comparison experiments are also insufficient. Aiming at these problems, this paper makes a comparative study on the application of yolov5 and fast RCNN in railway foreign object intrusion detection. First, the train and test dataset was established by image preprocessing, data cleaning and data labeling of UAV aerial images. Second, the canny edge detection algorithm combined with Hough transform was used to extract the track features for delineating the detection area. Finally, Yolov5 and fast RCNN, two widely used models, were used to train and test respectively based on our dataset for comparative studies. Experiment results show that YOLOv5 has better comprehensive performance in detection rate and detection speed, and Faster RCNN model cannot meet the requirements of real-time detection of track foreign objects intrusion.