Real-Time Train Detection Based on Improved Single Shot MultiBox Detector

Abstract

Autonomous environmental perception can be used in urban rail transit to extend the driver’s view. By installing cameras in the front of the train, the running line ahead can be detected. With the help of deep learning algorithm, drivers can identify trains at a long distance. However, complex environment makes the algorithm less robust to the identification of small targets. Therefore, the real-time and accuracy of front-train detection needs to be improved. In this paper, an improved Single Shot MultiBox Detector (SSD) algorithm is proposed. Compared with traditional image recognition method and original SSD, this method has higher accuracy in detection of the train, especially the small one. Moreover, it processes the images faster than traditional method. Experiments show that our method is very robust for the train detection in various illumination environment such as shadow, reflection, glare and high noise, and it reaches 95.38% mean average precision (mAP) and 26.3 frames per second (FPS) on our self-made dataset.