Research On Recognition Method Of Railway Operation Environment Abnormality Based On UAV Image And Laser Fusion

There are many factors that may endanger the normal operation of the train in the railway operation environment.If they cannot be found and handled in time,it may cause major safety accidents and bring great threat to the safety of people’s lives and property.The inspection of railway operation environment is the basis of anomaly recognition,but the inspection method based on manual inspection has some problems such as low efficiency,insufficient accuracy and insufficient inspection coverage when facing a wide-ranging and complex railway operating environment,and manual identification is prone to misjudgment.The existing inspection and anomaly recognition methods cannot meet the increasing demand of railway operation safety guarantee,and an intelligent and automatic inspection and anomaly recognition method for railway operating environment is urgently needed.Based on the above problems and requirements,this paper proposes a method for identifying abnormalities in railway operating environment based on UAV images and laser fusion.The main tasks are as follows:(1)Based on field investigation and theoretical analysis,two types of typical anomalies that have a great impact on railway operation safety are identified:construction and garbage accumulation,and the main hazards of these two types of anomalies and their own attributes o are analyzed in detail.On this basis,this paper proposes an anomaly detection method which combines the attributes of the anomaly object and the distance from the railway.(2)Build the inspection system of railway operation environment based on UAV and formulate a patrol inspection plan.The system integrates airborne visible light camera and lidar,and has the advantages of wide data acquisition range,high data accuracy,safety and efficiency.Based on the built UAV inspection system,the inspection scheme of railway operation environment is formulated.The scheme is committed to improve the inspection automation degree on the premise of ensuring safety,and has achieved good results.(3)Research on Airborne LIDAR point cloud data processing method.On the premise of analyzing the characteristics of the data,a series of classical algorithms are used to preprocess the point cloud data,including: outlier removal algorithm based on statistics,redundant point removal algorithm based on the nearest distance,ground point filtering algorithm based on cloth simulation filtering;then a large-scale point cloud semantic segmentation algorithm based on random sampling,feature aggregation and prototype fitting is proposed,which has achieved good results and overcomes the small sample problem of data to a certain content.This paper improves the point cloud clustering algorithm based on Euclidean distance,optimizes the distance threshold selection method,and proposes an irregular point cloud volume calculation method based on alpha shape algorithm,which has a great improvement compared with the existing algorithm;in the aspect of distance calculation between abnormal objects and railway boundary,this paper proposes a new method based on Alpha-shape algorithm.In the calculation of the distance between the object and the railway boundary,this paper converts this distance into the distance between the abnormal object and the railway center line to eliminate the error caused by incomplete scanning of the track edge.(4)Research on image data instance segmentation method.In this paper,firstly,two data enhancement methods are used to enhance the collected visible light image data of the railway operating environment,and transfer learning technology is used to solve the problem of insufficient general feature learning caused by insufficient data,and then an instance segmentation is carried out based on the YOLACT algorithm.The result can be used as a reference and supplement to the point cloud processing result.The experimental results show that the instance segmentation algorithm based on transfer learning and data enhancement has better recognition results.(5)Decision-level fusion method of point cloud data and image data.This paper proposes a serial data fusion method based on point cloud and image mapping.The method extracts potential anomalies based on the point cloud recognition results,and uses the image data recognition results to subdivide the potential anomalies to obtain the fusion recognition.The fusion recognition results combines the advantages of different data,and the recognition effect is better.