Real Time Detection Technology Of High Speed Railway Perimeter Based On Deep Learning

The intrusion of obstacles into railway perimeter poses a potential threat to the safety of railway operation system,so it is of great significance to get the accurate location of the railway perimeter and then judge whether the obstacles invade the railway perimeter to ensure the safety of railway operation.The real-time identification and alarm of different levels within the perimeter is conducive to timely taking different emergency measures.Most of the cameras installed along the railway are pan-head cameras,whose monitoring angle will change from time to time,so the position of perimeter in the monitoring image will also change,manually marking the perimeter can not meet the real-time requirements.So it is of great significance to study the railway perimeter detection algorithm which can automatically and quickly identify the perimeter in multiple scenes and has high accuracy.Traditional segmentation algorithm for perimeter detection slow speed and large amount of calculation,which means it cannot meet the actual demand.In order to solve the above problems,this paper proposes a perimeter detection algorithm based on deep learning,and clipped network to make the algorithm meet the real-time requirements.Firstly,this paper analyzes the feasibility of the existing deep learning object detection algorithm for railway perimeter detection.Although this method has the advantage of fast detection speed,it depends on the anchor box that can not adapt to the railway perimeter,so it is difficult to detect accurate result of perimeter.In order to solve this problem,this paper proposes Multi-box detection algorithm,which is based on the results of traditional object detection algorithm.Using Multi-box to identify the perimeter and combine the results,we can get a result that is more suitable for the perimeter than single box detection.However,there are still some problems,such as missing a small part of the area and unclear boundary between areas of deifferent levels.In order to solve the problem that the boundary of perimeter recognition based on the results of target detection algorithm is not clear,this paper proposes a scheme of perimeter recognition based on feature points,which abandons the anchor box of traditional object detection and retains the advantage of fast detection speed.;In order to verify the algorithm in the railway scene,the railway scene dataset is produced.Firstly,this paper collected the monitoring videos of multiple cameras along the railway,then established a rich railway scene sample library.Aiming at the task of identifying the boundary feature points,the dataset of heatmap is generated by annotation;Then,this paper improved the VGG16 convolution network model,and used multi-scale feature information to recognize the feature points of railway perimeter,and the railway perimeter recognition network model LDNet is obtained;Finally,the LDNet is trained by the heatmap dataset to get the optimal weight parameters.The trained LDNet achieves 96.11% of m PA and 92.75% of m Io U on the railway dataset,but the network still has the problems of large model and running speed not meeting the real-time requirements.In order to make the deep network model LDNet meet the actual needs of real-time perimeter recognition and small memory occupation,the network is compressed by network clipping.To solve the problem that the existing network clipping criteria can not directly reflect contribution to the output of kernel,a clipping criterion based on feature similarity is proposed,The contribution degree of convolution check output is defined according to the similarity between the output of LDNet and the output of each hidden layer.The network is compressed from 114.29 MB to 2.99 MB by a few times of clipping.After compression,m PA of the network increases by 0.26%,m Io U only decreases by 0.08%,the accuracy is almost has no loss,and the calculation time of single image is decreased by 64.9%.The perimeter recognition algorithm proposed by this paper can better adapt to the railway perimeter,and the recognition accuracy is significantly better than other algorithms.At the same time,it solves the problem of many parameters of deep network model and large memory occupation.Without GPU acceleration,it only takes 0.52 s to calculate a single image.