Research On Railway Small Target Detection Algorithm Based On Deep Learning

With the rapid development and popularization of high-speed railway,ensuring the safety of train operation has become more and more important.Machine vision technology based on deep learning has made new achievements in the field of target detection,which provides important theoretical support for railway foreign object detection.The existing target detection algorithms have the problem that the detection accuracy of small targets is not ideal in the task of foreign object detection in railway intrusion.In this thesis,the target detection algorithm based on deep learning is studied in the following aspects:(1)A SSD small target detection algorithm based on multi-scale feature jump fusion is proposed.Aiming at the problem of low detection accuracy of SSD algorithm for Railway Limited small targets,an efficient enhanced feature extraction method is proposed.The feature extraction performance of the network model is improved by fusing different size feature maps with the last layer feature map,so that the last layer feature map is rich in more details and semantic information.Secondly,in the feature fusion part,a new feature pyramid model is proposed to reverse the feature fusion of non adjacent deep feature maps,which not only makes efficient use of the rich semantic information of deep features,but also speeds up the detection speed of the network.The improved algorithm can effectively improve the performance of small target detection in railway intrusion data set.(2)A CenterNet occlusion small target detection algorithm based on density map module is proposed.Aiming at the missing detection problem of anchor frame overlap caused by occlusion in the target detection algorithm based on anchor frame,a global thermal map is proposed to efficiently respond to all key point information and obtain more location and quantity information of detection targets.Secondly,the elliptic Gaussian kernel function is used to replace the circular Gaussian kernel function for label coding,and the loss function including distance loss is introduced to obtain a high-quality prediction frame more suitable for the target.The improved detection model effectively reduces the missed detection rate of railway occluded target detection.(3)Make the railway violation data set,and jointly transfer and learn the alternating training network model.Firstly,the railway video surveillance and other image data are made into a new railway violation data set according to the VOC data set format;Secondly,the improved model is trained alternately by parameter based and instance based transfer learning methods.The improved model after transfer learning has good generalization ability in the task of Railway Limited small target detection.(4)Design and implementation of railway intrusion and foreign body monitoring system.In order to improve the detection accuracy of railway intrusion-limited small targets,this paper develops a railway intrusion-limited foreign object monitoring system platform by analyzing the functional and non-functional requirements of the foreign object detection system in the railway environment.The platform can not only monitor the railway environment in real time,but also have the function of detecting small targets and occluded targets in videos or images,which can meet the practical application requirements of railway intrusion monitoring of small targets.