Railway Obstacle Detection Algorithm Based On Deep Background Subtraction

Railway traffic safety is an eternal topic.The detection of foreign matter intrusion in railway lines has always been an important research topic in the field of rail transit.It has long-term research interests in theoretical research and practical engineering applications.Especially with the rapid development of Chinese railway construction and high-speed railway technology,there is an urgent need for a railway foreign object detection technology with high detection accuracy,good real-time performance,stability and reliability,which is of great significance in the field of railway security construction.At present,domestic and foreign scholars have proposed lots of mature programs in the field of foreign object detection,and have achieved certain practical application results.However,the detection method has much room for improvement in terms of accuracy,real-time and reliability.In recent years,deep learning and background subtraction,as two common visual detection algorithms,are widely used in the task of foreign object detection.But there are still some shortcomings.Deep learning has poor detection effect on untrained sample images,and background subtraction is easily interfered by environmental factors.Therefore,in this paper,with the intelligent recognition technology based on video surveillance,the image processing algorithm of deep background subtraction is studied,which achieves the detection of foreign objects in railway lines with good performance.Firstly,with the current research status and the overall demand of railway foreign object detection,an algorithm framework based on deep background subtraction and transfer learning is designed.Then,with the on-site railway video as the material,the sample images are intercepted and stored.12 typical camera scenes are also sorted out,and different foreign objects are manually labeled.What’s more,a database containing 40000 samples is established as the neural network training and test.At the same time,the block processing algorithm is designed,and the sample image is preprocessed by block downsampling.Different background models are also designed for the deep background subtraction algorithm,and the influence of different background extraction methods on the algorithm is verified.Next,based on the deep background subtraction algorithm,the network structure,network parameters and training methods of the pre-training model are designed.The neural network is trained and tested for different algorithm structures and image preprocessing methods.Under the scene data,99.6%of foreign object recognition accuracy is achieved.Finally,using the transfer learning method,the network is trained for each camera scene to improve the efficiency of training and identification for the algorithm model.The test results show that the foreign object recognition accuracy of the transfer algorithm can still reach 99%in a single camera scene.This paper combines deep learning methods with background subtraction through innovative ideas.By making up for the shortcomings between them,the deep background subtraction algorithm can solve the problem of accuracy,real-time and generalization for railway foreign object detection.It has powerful processing capabilities,overcomes the shortcomings of traditional detection algorithms,and has achieved good classification results in each group of experiments.