A Novel Strategy of Two-Stage Cascaded CNN and Overhaul Knowledge Distillation for Fast Railway Foreign Objects Intrusion Detection

The efficient and accurate detection of foreign objects invading railway tracks is of paramount importance in safeguarding the safety of train operations. Addressing the issue of the problem of the low efficiency of the existing foreign objects detection methods, this work proposes a fast railway foreign objects intrusion detection method based on cascaded convolution neural network and Overhaul Knowledge Distillation. First, a two-stage cascaded convolution neural network is built.The first stage can identify whether the railway images are intruded by foreign objects or not. This is achieved by a light weight image classification network. The use of lightweight classification network can reduce the use of the object detection network, thus improving the overall efficiency of the railway foreign objects intrusion detection method in this paper. Secondly, this paper employs the Overhaul Knowledge Distillation algorithm to train a lightweight network that is supervised by a larger network, so that the lightweight network constructed in this paper also has satisfactory image classification performance. Finally, the YOLOv3 object detection network is used to detect the foreign object image classified by the first level network. The experimental results demonstrate that the accuracy of the image classification network proposed in this paper is competitive to the classical backbone network, and the FPS is about 50-70 higher than the comparison method.