A deep-learning-based real-time object detection technique in railway transportation

Abstract

In computer vision, object detection is the process of finding or identifying things in pictures or videos. It provides a quick and effective method for classifying objects. Its extensive applications in object recognition, separation, and detection have advanced safety and convenience in our contemporary life. In this paper, we introduce a model that detects the object in real time at railway crossings. In the experiments, I used YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x of the YOLOv5 family. Each model was trained separately to see which one performed better in terms of recall, mAP@0.5, mAP@0.5:0.95, and precision. To train and validate the object detection system, I used open-source images captured on railway crossings and also used online data. The dataset consists of 949 images containing eight classes, almost 550 are raw data which is captured on railway crossings and almost 400 images are collected online. The dataset I used divided includes 758 images for training and 189 images for validation. After the experiment of YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x model, YOLOv5x obtained the highest average of precision at 89.4%, recall at 85.5%, and mAP at 95.6%. So YOLOv5x is the most stable method followed by YOLOv5s, YOLOv5m, and YOLOv5l. The experimental results that YOLOv5x algorithms are accurately detecting real-time objects in railway crossings. It will be a great help in reducing the collisions of trains and also reducing the accident rate by real-time object detection.