| With the rapid development of high-speed railway construction,although the casualties caused by railway traffic accidents have decreased slightly with the improvement of detection technology and protection technology,they remain high.It is an urgent problem to effectively identify and avoid missed detection and false detection of intrusive foreign objects that intrude into the railway safety limit and affect the railway traffic safety and the normal operation of the railway system.The railway foreign body intrusion limit detection algorithm based on deep learning can replace the manual inspection and traditional detection methods,and complete the effective division of the intrusion limit in the field of railway foreign body intrusion and the effective detection of the foreign object to be detected.However,there are also phenomena such as insufficient railway data sets,inaccurate division of railway safety boundaries,excessive detection model parameters,long detection time,and missed detection and false detection.In view of the existing shortcomings,this paper firstly constructs a high-speed rail boundary data set for railway intrusion limit area division and a railway foreign object detection data set containing foreign objects for high-speed rail foreign object intrusion limit detection.Secondly,in order to accurately divide the dangerous area and the safe area of the railway,the image semantic segmentation algorithm based on the fully convolutional network is adapted to the fully convolutional network.(1)The use of depthwise separable convolution structure instead of standard convolution greatly reduces the number of parameters and calculations in the convolution operation,and enhances the feature expression ability of the network through the connection of the reverse residual structure to solve the problem of training time.The gradient disappearance problem occurs when the number of network layers increases.(2)Introduce the empty space pyramid structure to increase the perceived size of the network and improve the network feature extraction ability without reducing the resolution.(3)The low-level and high-level feature information of the image is fused to retain more detailed information in the image,thereby improving the accuracy of the division of railway encroachment boundaries.Finally,in order to improve the accuracy and speed of detection of high-speed iron foreign objects intrusion limit targets,and avoid missed detection and false detection of intrusion-limited foreign objects,the YOLOv4 model is adaptively improved.(1)In view of the problem that the model is too large,on the one hand,Mobile Net V3 is used to replace the original feature extraction structure,and the Inception-Res Net-v1 structure is introduced to improve the problem of lower detection accuracy due to the reduction of model parameters.(2)Optimize the output of the network to improve the phenomenon of missed detection and false detection of the small target to be detected by the network,thereby improving the overall performance of the model.The improved railway intrusion limit image semantic segmentation algorithm and the railway foreign body intrusion target detection model have been greatly improved in accuracy and speed,which are of great significance to the realization of efficient,intelligent,and automated development of railway safety system safety detection.meaning. |
