High-Speed Railway Clearance Intrusion Detection Using Feature Fusion Enhancement Neural Network

With the rapid development of China’s railway and the continuous improvement of operating mileage,the operation safety of high-speed railway has been paid more and more attention.However,due to the imperfection of the basic protection measures along the railway,foreign body intrusion often occurs,which seriously threatens the national property and people’s lives.The existing foreign body intrusion detection methods have defects in the installation and maintenance cost,anti-interference ability or detection accuracy.Therefore,this paper aims to propose a more accurate,fast and stable railway clearance intrusion detection method based on deep learning theory,so as to timely predict the potential safety hazards and improve the safety level of train operation.In the railway scene,intrusive foreign objects are mostly small targets,but the existing deep learning algorithms have insufficient detection capabilities for small targets,and cannot meet the requirements of foreign object intrusion detection in the railway field.To solve this problem,this paper proposes a feature fusion enhanced network(FFE-Net)architecture,which includes three parts: anchor filter module(AFM),feature fusion enhanced module(FFE)and refinement detection module(RDM).Among them,FFE can effectively enhance the detection effect of small targets and targets with large aspect ratio,while AFM and RDM solve the problem of imbalance of positive and negative samples in the network training process through cascade regression.In addition to the network architecture,the training strategy also has an important impact on the detection performance of FFE-Net.The improvement of training strategy in this paper mainly includes two parts: firstly,aiming at the shortcomings of traditional anchor matching algorithm in positive and negative sample classification,a high-quality anchor selection strategy(HASS)is proposed,which is more in line with the characteristics of FFE-Net cascade regression,and provides a new standard for positive and negative sample classification of anchors;secondly,in order to improve the positioning accuracy and training stability of the network,a weighted fusion localization loss(WFLL)based on maximum likelihood estimation is designed by fusing DIo U Loss and Smooth L1 Loss.This method provides a new idea for weighted fusion of multiple loss functions in target detection network training.Since FFE-Net is a detection algorithm for single video frame,it will cause repeated alarms when detecting consecutive video frames.In order to solve this problem,this paper designs a repeated alarm filtering algorithm based on Deep Sort framework.The algorithm can determine whether it is the same target by calculating the correlation of detected objects in different frames,and then decide whether to send new alarm.By collecting the video images of the railway scene,this paper constructs the railway foreign body intrusion sample database including the empty scene image,pedestrian intrusion image and train entry image.The experimental results on this dataset show that when the input image resolution is 320 × 320,FFE-Net achieves the detection performance of 88.55% map and 26 FPS on Geforce GTX Titan X,which can meet the application needs of railway foreign body intrusion detection.And the repeated alarm filtering algorithm is also verified by pedestrian intrusion video,showing good real-time performance and robustness,which can meet the actual use needs of railway field.