Research On Detection And Tracking Algorithm Of Railway Foreign Objects In Infrared Weak Light Environment

As one of the important infrastructures of the country,the railway system undertakes the important tasks of people’s travel and cargo transportation.The railway system plays an irreplaceable role in the economic development of the country.However,with the increase of railway mileage,the environment of train operating also has become complicate and variety.Foreign body intrusion accidents frequently occurs in railway environments.Railway foreign objects refer to pedestrians,vehicles,rocks and other objects which intrude into the railway clearance and endangers the normal train operations.Railway foreign objects pose a serious threat to the safety of railway operations.Therefore,the detection and tracking of foreign matters in railway plays an important role in ensuring the safety of people’s property and railway operations.With the all-weather operation of China’s railways,the railway operating environment in night has become more complicated.Target detection based on visible light is easily affected by intensity of illumination and cannot be effectively imaged at night.Infrared imaging formed by infrared light which reflexed back from objects.Therefore,infrared imaging is used for night traffic monitoring and detection.Infrared imaging use the receiver to sense the infrared light which reflected from the environment.In the low-light railway traffic environment at evening,infrared imaging has difficultly of poor definition,color loss,and less texture details.In addition,the railway infrared imaging at evening is interfered by the noise in the railway environment.The research on detection and tracking algorithm of railway foreign objects in infrared weak light environment.It remedy the shortcomings of the visible light imaging.It also provides technical guarantee for the safety of railway operations which in all-weather environment.This dissertation summarizes the research of railway foreign object detection and tracking algorithms at domestic and abroad.Then,the dissertation analysis the deep learning infrared foreign object detection and tracking algorithms.The main contents include the following three directions: Detection of object intrusion under infrared low light condition based on multi-feature and attention enhancement network.Lightweight detection of railway object intrusion based on spectral pooling and shuffled-CBAM enhancement.Infrared railway foreign objects tracking based on spatial location and feature generalization enhancement.The main research and contributions are as follows:(1)Aiming at the problems of insufficient feature extraction and low detection accuracy of existing deep learning algorithms based on infrared targets in weak light environments,this chapter proposes an anchor-free foreign object detection algorithm based on multiple feature fusion and attention enhancement.Firstly,the Center Net model is improved by multi-scale adaptive feature fusion.Secondly,the Dilated-CBAM module is introduced to extract features,which improves the detection precision of the anchor-free network.Finally,using the Smooth L1 loss function to improve the convergence speed and stability of the network training process.Experimental results verify that the proposed method in this chapter has an average detection accuracy improvement of 8.03% compared to the Center Net method.The proposed method has excellent detection real-time performance.(2)Aiming at the difficulty of large network parameters and low real-time detection efficiency for railway foreign object intrusion detection algorithms in infrared weak light environments,a lightweight detection method of railway foreign object based on spectral pooling and shuffled-CBAM enhancement is proposed.Firstly,the deep separable convolution is used to reduce the parameter amount of the Darknet53 network to achieve rapid feature extraction of infrared railway foreign objects.Secondly,the spectrum pooling enhancement module with semantically guided is used to improve the feature quality of infrared image down-sampling.Then,we propose the shuffled-CBAM module to extract the key features of the infrared target,The shuffled-CBAM module improve the accuracy of infrared target detection in the network.Finally,the anchor-free lightweight network is used to detect the railway foreign body intrusion.The lightweight network overcomes the disadvantage as the poor real-time performance of the non-maximum suppression operation which happened in anchor frame detection algorithms.The anchor-free lightweight network reduces the amount of computation while improving the detection efficiency.Experimental results show that the proposed lightweight method improves the detection rate to 39 frames per second,and detect infrared railway foreign objects quickly and accurately.(3)Aiming at the problems of low tracking accuracy caused by infrared target position movement,morphological changes,and feature occlusion in existing deep learning algorithms for railway foreign object tracking,a feature enhanced railway foreign object tracking algorithm is proposed.Firstly,the multi-scale cascade Ghost Net network is used to enhance the ability of the model to extract target features.Secondly,the infrared feature is enhanced by spatial location and feature generalization module.This module enhanced the foreign object spatial location and generalization morphology.Then,the pyramid prediction network is used to obtain the detection anchor frame,classes and confidence of infrared railway foreign objects.Finally,the Deep SORT tracking algorithm with improved category and confidence display is used to realize the tracking of railway foreign objects in infrared weak light environment.Experimental results verify that the tracking accuracy of the proposed method reaches 73.9% and the average tracking rate reaches 11.3 frames per second.This method has excellent infrared target tracking performance.