Automatic Counting of Railway Tools Based on Deep Learning.

The daily railway operation and maintenance suffers from a severe problem of the loss of railway tools. Aiming at the problem, an automatic counting method of railway tools based on deep learning is proposed. Our method is based on toolkit images obtained before a toolkit is delivered to a worker and after it is returned. By recognizing and comparing railway tools in toolkit images, our approach can detect missing tools automatically. Our work extends the research of object detection to the practical application of railway operation and maintenance. In order to resolve the sparsity problem of image samples, an image dataset augmentation algorithm is used for oversampling. Combined with the transfer learning strategy, our approach is able to count railway tools based on images automatically and accurately in complex outdoor environment. Experiments were conducted based on real-world datasets. Results show that our method can detect railway tools accurately with a mAP of 83%, which satisfies requirements of practical applications. Above all, our work provides a strong technical basis for intelligent railway operation and maintenance.