Encoder Teacher-student Network for Foreign Object Intrusion

Foreign body intrusion in high-speed railway is often accompanied by the heavy loss of life and property. Therefore, the monitoring of foreign objects in the railway industry is of great significance to maintain the safe, reliable and stable operation of the entire railway. The continuous development of deep learning theory and the improvement of computation resource provide new possibilities for the identification and judgment of foreign object intrusion by using the existing cameras. In this paper, we propose an improving Teacher-Student network based on knowledge distillation, which is able to judge and identify foreign body intrusion through 2K video stream captured by the railway, and propose an attention mechanism based on probability density to effectively improve the model identification ability due to the needs of the railway scene. Compared with the supervised learning approach, this paper utilizes the unsupervised learning approach to effectively cope with the irregularity, complexity and even unknowability of foreign objects. Experiments on the private dataset prove that the AUC value can be as high as 0.99, the false positive rate and the false negative rate are less than 2%, and the real-time performance can meet the practical needs of the railway. Meanwhile, our model provides an alternative approach for foreign body intrusion.