Visual Monitoring of Railway Obstacles using Deep Learning-Based Methods

One significant aspect contributing to the overall improvement of railway operations and safety is the detection of obstacles along the tracks. This advancement not only enhances the efficiency of rail transport but also ensures greater accuracy and reliability. Thanks to recent advancements in artificial intelligence technologies, obstacle detection, and self-driving methods have witnessed remarkable progress. In this project, a structure has been designed that enables the real-time detection of obstacles on the train track with a balance between accuracy and volume of calculations. The proposed method involves identifying the train path and checking for the presence of obstacles in this area. Three segmentation models, PAN+SE-ResNet, PAN+EfficientNet, and PAN+NF-Net, were designed, trained, and validated by a set of color images. Results show that the PAN+NF-Net rail segmentation model is more accurate than the other two models with a few tenths of a percent difference. Meanwhile, the computing load of the PAN+EfficientNet model is about half and one-fifth of the computational load of the PAN+NF-Net and PAN+SE-ResNet models, and it offers faster processing speed. To design the object recognition model, the fifth and seventh versions of the YOLO algorithm were trained by a dataset. The results show that the accuracy of the YOLOV7 model in detecting obstacles is 17.6 percent higher than the accuracy of the YOLOV5 model. Finally, obstacle detection Operation was accomplished by checking for overlaps between the mask of the train's path and the bounding boxes predicted by the object detection model.