Research On Recognition And Classification Of Foreign Objects In High-speed Railway Lines Based On Deep Learning

The intrusion of foreign objects will seriously threaten the safety of trains during the operation of high-speed rail lines.At present,the main body of railway monitoring system in China still relies on manual monitoring and inspection,which not only consumes a lot of manpower and material resources,but also has the disadvantages of low efficiency and long detection time.However,the traditional target detection algorithm applied to high-speed rail foreign body detection also has the problem of difficult feature extraction,and the extracted features are difficult to fit the complex background environment of the high-speed rail scene.In recent years,the target detection method based on deep learning has been developed rapidly.The application of deep learning to the problem of railway foreign body intrusion detection is a new direction.The deep learning methods was used to study foreign body identification and classification of high-speed railway lines.During the research process,the Guiyang high-speed railway station monitoring video data is used to make high-speed railway foreign body identification and classification data sets.In order to solve the problem of small samples of special weather scenes and night scenes in the data set,the sample style of high-speed rail scenes transformed by Cycle GAN network to obtain new samples which could supplement the data set,and select suitable pictures from other data sets to complete the high-speed rail foreign body recognition and classification.Then,this data was used to set to experiment with several excellent target detection algorithms,compare the detection speed and detection accuracy of each network,and select the YOLOv3 network with the best comprehensive detection performance as the basic network for high-speed rail foreign body recognition and classification research.Finally,according to the characteristics of the high-speed rail foreign body recognition and classification data set,the YOLOv3 network is improved.First,the switchable cavity convolution SAC is used to replace the first four 3×3 convolutions in the feature extraction network,and the detection accuracy of the model for large-scale targets is improved by increasing the receptive field of the convolution.Then,in order to improve the model's detection accuracy of small objects and the FPN structure of YOLOv3,thesecond down-sampling output in the YOLOv3 feature extraction network was integrated to the feature map.The output scale is a new prediction special layer of104×104,and the new FPN structure is integrated more shallow information from the backbone network which strengthens the network's ability to detect small targets.In this paper,the improved YOLOv3 ablation experiment and the comparison experiment between the improved YOLOv3 network and other excellent target detection networks prove the effectiveness and correctness of the high-speed rail foreign body recognition and classification network based on the improved YOLOv3.Experimental results show that the improved YOLOv3 high-speed rail foreign body recognition and classification model is compared with the original YOLOv3 network.When the detection speed is slightly reduced,the average detection accuracy reaches79.1%,which is 4.3% higher than the original network.At the same time,the improved YOLOv3 high-speed rail foreign body detection network has better detection accuracy and real-time performance than other target detection networks.