Research On 3D Imaging And Point Cloud Segmentation For Railway Environment Monitoring

The environmental monitoring of railway is a significant means to ensure the security of train operation and improve the quality of railway operation.Compared with traditional manual inspection,the detection technology based on machine vision has significant advantages,such as high efficiency,high accuracy,no damage and realtime recording.Compared with the detection technology based on 2D data,the detection technology based on 3D data has higher reliability,accuracy,and better application prospect,and thus it has become an important research content of railway operation environment monitoring.Aiming to the state detection requirements of fastener and power line facilities in the integrated intelligent inspection system of railway environment,this study proposes that the three-dimensional imaging technology for railway tracks and the point cloud segmentation method for fastener and power line,respectively.The findings have several following contributions:First,Design and calibrate the railroad 3D linear structured light scanning module.In terms of the comprehensive inspection system in a larger scope of the railroad of the3 D data acquisition task,we choose the simple structure,high precision,non-contact linear structured light imaging technology as a solution of imaging,in combination with the equipment selection principle of machine vision application,we choose camera,lens,laser to build an scanning platform which scanning length is 1.5 m.Because the camera in the line-structured light scanning platform can only photograph the object illuminated by the laser line,it is not possible to use the chess board to calibrate the camera’s intrinsic parameters and lens distortion in advance.The mathematical model of the imaging is improved by combining the conditions that the pixels of CMOS are square and the rows and cols of pixels are perpendicular to each other.In this paper,we proposed a calibration method,which firstly uses line target to calibrate the image distortion and uses a toothed target to calibrate the homography between the image plane and the object plane.The calibration experiments showed that the re-projection error of calibration was less than 0.1 pixel.Accuracy verification experiments showed that the calibration method can achieve the scanning accuracy of0.5mm,which can meet the application requirements of railroad scanning.Second,a point cloud segmentation method for the fastener region was proposed for the point cloud of railroad scan.More specifically,firstly,the crossover method of horizontal and vertical gray threshold was used to locate the fastener region from the whole point cloud.Then,existing research has shown that the point cloud segmentation algorithm is susceptible the problems of over-segmentation and under-segmentation,and thus,we proposed an super-voxel and graph based fastener region segmentation method—namely,the first super-voxelization the fastener cloud into border with good performance,and adopted the super voxel aggregation method based on local adjacency graph form over-segmentation to segments.Finally,there existed an overlapping region segments merged to form the final segment result.Experimental results showed that the proposed algorithm has certain advantages in the performance of fastener segmentation and anti-noise,in comparison with the existing algorithms.Thirdly,a power line segmentation method based on 3D point cloud in railway scene is proposed.In comparison with the existing power line segmentation methods,this method has the characteristics of less computation and less computation time with the large point cloud scale.It can be divided into two parts: power line regional location and power line detection.Firstly,the digital elevation model was constructed with an accuracy of 0.2m,then,the railroad region was accurately located by image binarization,contour search,convex boundary extraction and optimization,and the power line region is obtained by ground point filtering according to the elevation difference.Secondly,the progressive probabilistic Hough transform line segment detection was carried out in the power line area to obtain multiple candidate line segments.According to the direction distribution of power lines and the number of neighbor points,the false detection results were excluded and then were converted back to 3D point cloud.Finally,the algorithm was run twice to segment the second layer of power line point cloud.Qualitative and quantitative analysis of the results of power line segmentation showed the rapidity and robustness of the proposed method.