Railway Obstacle Detection Algorithm Based On Deep Neural Network

In the field of rail transport, the real-time and accurate intrusion detection is always an important topic, owning lasting research heat in the field of practice and research. Especially with the deepening of the high-speed railway technology and the development of China’s high-speed railway construction, designing a detection algorithm with excellent recognition effect to be applied to the actual operation of the railway has a more important significance.Currently, traditional foreign body detection algorithms, which are based on video technology, are depended on background frame difference method. They are easily influenced by the scene and light changes. The false detection rate is higher and they can’t satisfy the need of the railway site long-term on-line detection. Deep neural network, as a newly emerging machine learning algorithm, has more complex network structure, better feature extraction and the network training mode compared with the traditional artificial neural network, owning a powerful processing capability in the field of image.In this paper, whether the railway line is occupied by trains is selected as the research task and the detection algorithm based on deep neural network is designed. Then the optimization methods and generalization performance are verified with the training sample set of actual railway scene images. First of all, on the basis of the collected video of the railway, the automatic classification algorithm is designed with the traditional detection method and a quantitative and accurate classification of the image database is constructed with the manual checking. Subsequently, five-layer deep belief networks are designed to realize the image recognition function and the structure and parameters optimization methods are studied by single camera images to realize the expected recognition task. Finally, the generalization performance of the algorithm is tested using of different camera images, in order to verify that the algorithm has good generalization performance by adjusting the training samples and the network structure.By using the deep neural network, this paper changes the original inherent foreign body detection mode, replaced with the method of scene image classification, to avoid various errors and shortcomings of the detection and extraction method. Good effect has been achieved in the actual scene video test and this shows that the algorithm has good recognition ability and practical significance.