Real time train, vehicle and pedestrian detection system for automated railway level crossing

Abstract

In this project, we have analyzedimages and videos ofthe train from different angles in different train stations of Dhaka in different environmental conditions. With these recorded videoswe have developed a real time train detection system. We have usedOpenCV, SSD model, and frozeninferencegraph for increased detection speed. In addition, the system can perform dynamic detection of vehicles and pedestrians. In this detection system, we compare the real time image with the model images containingfor vehicle detection with preset threshold values. If the resultant value is higher than the threshold value, then the system detects the presence of the object i.e., vehicles or pedestrians.In the beginning we worked with images to make this detection system. After being successful there, we shifted to the moving images i.e.,video. After successfully analyzing the images from the video, we started working with real-time video footages. In the case of real-time videos, many frames come together, and they are converted into images, and are matchedwith the model of the detection system. With higher frame rates, the system performs slower. That is why we optimized the frame rate so that the system can operate in real time with adequate accuracy. Although we started working withYOLOmodel,later we shifted to theSingle Shot Detector (SSD)model as the later was faster and more accurate. The proposed system can enable an automated level-crossing system in Bangladesh for which accurate and faster detection of the train is a key component.