Reserach On Detection Algorithm Of Railway Foreign Body Intrusion Limit Based On Deep Learning

The detection of foreign body intrusion is of great significance to ensure the safe operation of railway trains.The early detection of foreign objects mainly relied on manual inspection,but the intrusion of foreign objects is very sudden,which makes it impossible to detect foreign objects in time only by relying on manual inspection..Although the detection method based on traditional machine vision can break the limitations of manual inspection and realize intelligent detection,the artificial features used in this method are easily disturbed by abnormal phenomena such as occlusion and strong light in the detection environment,which may cause serious problems.Therefore,it cannot adapt to the complex operating environment of the railway.The deep learning method has the advantages of high detection accuracy and strong robustness,and can adapt to various detection scenarios.Therefore,this paper adopts the detection method based on deep learning to detect the intrusion of railway foreign bodies.The main work can be divided into the following aspects:Firstly,in order to solve the shortcomings of target detection based on artificial feature extraction operators,it is proposed to apply the YOLOv3 target detection algorithm in deep learning to the detection of railway foreign body intrusion,and to improve the shortcomings of the original algorithm.First of all,in view of the characteristics that deep learning algorithms usually consume a lot of computing resources,an intra-layer multi-scale residual block(MRblock)based on a new convolution calculation method is designed,and this residual block is used to replace the residual block in the original YOLOv3 network,this module changes the original single-channel convolution path into a two-channel convolution path based on image frequency,and reduces the overall computational complexity of the algorithm by downsampling the low-frequency feature map;secondly,the spatial pyramid module and the adaptive spatial feature adaptive fusion module increases the network’s ability to perceive semantic information of different scales,and reduces the semantic information conflict problem that occurs when the feature pyramid fuses features of different scales.The superiority of the improved YOLOv3 algorithm in detection accuracy,computational complexity and detection speed is verified by testing on the self-made railway foreign body data set.Secondly,in view of the lack of identification of the intrusion limit of railway foreign objects in the existing foreign object detection research,the position information of railway boundary is obtained by identifying railway track first and then expanding the coordinate position of railway line according to the definition of building boundary under standard gauge.Using the segmentation algorithm based on the line anchor frame in the detection of the track line,and using the preset line anchor box to identify the position of the track line can significantly reduce the calculation amount of the segmentation algorithm and improve the calculation speed.It can be proved that the detection algorithm in this paper has higher detection accuracy and faster detection speed through the comparative experiments on the track line data set.Thirdly,a railway foreign body intrusion detection system is built by combining the improved YOLOv3 detection algorithm and the track identification algorithm based on the row anchor frame.According to the previous experimental results on the target detection network and the track detection network,the optimal network parameters are selected.By testing the detection performance of the system on the foreign body intrusion data set,it can be proved that the foreign body detection model presented in this paper can accurately judge whether objects constitute intrusion in real time.The railway foreign body detection system proposed in this paper integrates two functions of foreign body detection and railway limit detection.Compared with other foreign body detection methods using only target detection algorithm,it can effectively reduce the problem of misdetection caused by the detection of objects outside the limit as foreign body intrusion,so it has higher practical value on site.