Research On The Application Of FCN And CNN In Railway Intruding Objects Recognition

With the rapid development of high-speed railway,the safety of railway transportation is becoming increasingly important.However,the invasion of foreign objects may pose a great threat to the safe operation of railway,seriously endangering state property and personal safety.Now,in order to solve it,there is a technology detecting the invasion clearance of railway foreign objects required with high precision,good real time and stability,which plays a significant role in the safe operation of railway.At present,the application of the deep learning algorithm in the field of image detection and recognition provides the possibility for the recognition of the railway intruding objects,so the thesis is aimed at developing a set of algorithm detecting the invasion of railway foreign objects,which is based on deep learning and can recognize as well as detect foreign objects invading railway.Thus,it can provide certain guarantee for the safe operation of railway.To accurately recognize and detect the foreign objects invading railway,to begin with,the thesis collects many on-site surveillance videos to create a railway database,mainly including an empty scene,railway image with operation train and railway image with pedestrian intrusion.Besides,based on fully convolution network,building sub-network aggregation and sub-phase aggregation to make down-sampling on the basis of improved Xception network in the encoder stage,and use bilinear interpolation and feature fusion method in the decoder stage to make up-sampling to divide rails quickly and accurately,to determine the limits of railway.At last,to quickly and accurately detect and recognize the target,the thesis is based on SSD network and introduces the number of the balanced positive and negative samples of the loss function Focal Loss on the basis of original network.Meanwhile,introducing the loss function DIOU Loss helps the network astringe better and faster.To solve the problem of less effective detection of small targets,the thesis introduces deconvolution and feature fusion to strengthen the effect and use convolution kernel L1 norm to cut and compress the network,thus,the computational complexity and its time can reduce,which may truly achieve the balance between the accuracy and its real time.Railway dataset and PASCAL VOC suggest that the algorithms in the thesis have good universality and robustness,which meets the demand of practical use on the railway site.In this paper,the fast semantic segmentation network based on the improved Xception network is used to divide railway boundary area and finally makes the accuracy(IOU)reaches 95.2%,and detection time of a picture is 3.85 ms.After that,the improved network SSD has been used to recognize targets,as a result,the accuracy is entirely 99.95% and the detection time of a picture is 31 ms,which is more practical and real-time,and it is also very important to the safe operation of railway.