Research On Automatic Identification Technology Of High-speed Train Orbital Intrusions

With the rapid development of high-speed railway in China,high-speed trains has become one of the important means of transportation for people’s daily life.The operational safety of high-speed trains has become particularly important to ensure the safety of people’s lives and property and to reduce the loss of national assets.Research has found that the orbital intrusion is one of the main factors affecting the operational safety of high-speed train.Therefore,it is of great importance to detect and identify orbital intrusions timely and effectively.At present,the detection of orbital intrusions based on object detection algorithm has gradually become the mainstream research direction.According to the characteristics of orbital intrusions,this paper mainly studies the application of the target detection algorithm in orbital invasions detection and recognition.The main research contents are as follows:(1)Establish orbital invasions datasets.Establish orbital intrusion datasets according to the feature of the algorithms of target detection that it requires a large amount of image data onto training.The dataset comes from self-collected images,networks,and public datasets.Adopt for enriching the dataset,prevent over-fitting,and improve the generalization of the model,the image was cropped,flipped,added noise,and adjusted brightness and contrast scale.(2)Dividing the orbital boundaries.It is found that the method of dividing the orbital boundary based on track cannot be the basis of judging whether the high-speed train tracked to invade the track.As long as it enters the protection network of the high-speed railway,it will affect the driving safety of the high-speed train and should be regarded as the orbital invader.Therefore,this paper takes the protection network of the high-speed railway as the orbital boundary.(3)FCOS implements orbital invasion detection.The FCOS network detects the target by pixel-by-pixel multi-scale prediction;the ResNet network is used to extract features,and the FPN network is used for feature fusion.The multi-scale detection head calculates the classification loss,regression loss and centerness loss for each output of the FPN.Through experimental verification,FCOS has a good detection effect on the COCO dataset.In order to improve the detection accuracy of orbital intrusions,this paper has made the following improvements:(1)Res2Net50,which represents multi-scale features at the granularity level,is introduced as the backbone network to enrich the receptive field of each network layer;(2)the C2 layer output of Res2Net50 is introduced as an input of FPN,and more feature fusion is adopted to increase the complexity of FPN network and the receptive field of each layer output feature map;(3)A new detection head is added,which is obtained by upsampling the P5 layer of FPN.(4)Introduce DIoU loss as the regression loss function.The mAP of the improved FCOS network on the self-made VoC datasets reaches 93.71%,which is a great improvement compared with the original FCOS network.(4)Build a real-time detection model.The imported video stream is preprocessed by OpenCV open source computer vision library.Comparing the noise reduction effect of mean filtering,median filtering,Gaussian filtering and bilateral filtering on the image,the median filtering algorithm is selected as the denoising algorithm;the adaptive histogram equalization algorithm is used to adjust the contrast of the image.The pre-trained FCOS model is loaded by OpenCV to detect the orbital intrusion in real time.