Railway Obstacle Detection Algorithm Based On Transfer Learning

Because of human factors or natural disasters,the phenomenon of foreign objects intruding into the railway limit has occurred sometimes.It is a potential threat to the safety of railway operation.Therefore,it is of great significance to monitor the security of the railway clearance in real time and it is important to achieve the early-warning,alarm and classification of foreign object intrusion.With the rapid development of high speed railway in China,it is urgent for the technology of foreign body detection in railway with high recognition accuracy,good performance in real time and cost-effective.Aiming at the existing integrated surveillance videos in high-speed railway,this paper studies on the algorithm which can process videos fast based on deep neural network,and achieves the optimization of processing algorithm and the accurate detection of foreign objects in different monitoring scenarios by using transfer learning.Based on the requirements of foreign objects' detection in high-speed railway,a framework of detection algorithm based on deep belief network and transfer learning is designed in the paper.Then it takes full advantage of video resources taken by cameras along the railway and images need be collected and stored from the video resources.After all the former work is done,typical samples of trains,foreign objects,weather,light and so on need to be sorted out.Following that,the images need be pre-processed by the methods of annotation,grayscale and down sampling.At last a mass image database containing 80 thousand samples is created as a sample set for training and testing of deep neural networks.Then,we design the structure and parameters of the pre-trained model based on the deep belief network,and finally establish a model with the structure of five levels including an image input layer,three feature extraction layers and a classification recognition layer.Using a large training sample set of a single camera to train this pre-trained model,the accuracy of foreign object recognition can reach 99%.Then,we use the sample set to train the pre-trained model,and select the parameters and structure that can be transferred by contrastive experiments.And the chosen parameters which have good effect on transfer learning are transferred to the other cameras' training and detecting to achieve real-time monitoring of the whole cameras along railways.The results of the test show that the accuracy rate of the foreign object recognition can still reach 99%by using the small training samples.This paper will study the method of transfer learning which is introduced into the monitoring work against foreign bodies in railways.Transfer learning can be used to solve new problems by the ability of areas' or knowledge's transfer to handle problems about slow detection speed and poor generalization performance caused by changes of scenes and cameras' cooperation in the video monitoring.Finally,we can extract more essential features using the deep learning algorithm to make recognition rate higher.