Classification of Metro Facilities with Deep Neural Networks

Abstract EN

Metro barrier-detection has been one of the most popular research fields.

How to detect obstacles quickly and accurately during metro operation is the key issue in the study of automatic train operation.

Intelligent monitoring systems based on computer vision not only complete safeguarding tasks efficiently but also save a great deal of human labor.

Deep convolutional neural networks (DCNNs) are the most state-of-the-art technology in computer vision tasks.

In this paper, we evaluated the effectiveness in classifying the common facility images in metro tunnels based on Google’s Inception V3 DCNN.

The model requires fewer computational resources.

The number of parameters and the computational complexity are much smaller than similar DCNNs.

We changed its architecture (the last softmax layer and the auxiliary classifier) and used transfer learning technology to retrain the common facility images in the metro tunnel.

We use mean average precision (mAP) as the metric for performance evaluation.

The results indicate that our recognition model achieved 90.81% mAP.

Compared with the existing method, this method is a considerable improvement.