Track Extraction And Railroad Object Segmentation Using MLS Data

With the rapid development of automatic driving industry,multi-sensor fusion technology is gradually emerging.Multi-sensor data fusion,such as GNSS / INS /lidar,is the core technology of three-dimensional reconstruction and environmental perception,which is widely used in the fields of automatic driving and facility inspection.Taking the railway environment as an example,the 13 th five year plan of the railway pointed out that it is necessary to accelerate the application of high-precision technologies such as Beidou in the railway environment and realize the intelligent and modern development of the railway.Railway intelligent means that the sensor platform not only needs to clarify its own position,but also needs to identify and extract the ground objects under specific scenes,so as to meet the high-level perception and planning control tasks.With the construction of Beidou foundation enhancement network and China’s mainland environmental monitoring network,the nationwide datum station data makes it possible for the seamless and precise positioning of the whole country.In recent years,the cost of inertial measurement unit has been decreasing.The combination of MEMS MEMS MEMS level inertial navigation data and GNSS can meet the needs of precision positioning in most cases.In the above context,the airborne platform composed of sensors such as GNSS / INS / lidar makes it possible to locate and sense the railway environment.This paper focuses on the construction of railway geographic information database and the specific needs of railway inspection,aiming to use the above sensors to build a three-dimensional database of railway corridor.On this basis,the railway track recognition and topology reconstruction,as well as the segmentation and extraction of the surrounding railway objects are implemented.The main research work and phased results are summarized as follows:(1)considering the characteristics of sensors and the requirements of railway patrol inspection,a set of hardware platform for multi-sensor data acquisition and processing is built and calibrated.The platform can achieve the centimeter level positioning accuracy under dynamic conditions,including horizontal positioning accuracy 3cm and elevation positioning accuracy 5cm.Considering the distortion effect of point cloud in high-speed environment,the distortion of sensor platform can be compensated by the estimated speed,which can meet the needs of real-time large area 3D reconstruction.(2)based on the data of Beidou foundation enhancement system,a set of fast extraction algorithm of railway track is proposed.The test results show that when the LiDAR frame rate is 10 Hz,it can be carried out in real time on the low-cost processing platform.Compared with the manual annotation data,the precision / accuracy /sensitivity of the extracted track points are 99.68%,97.55%and 66.55%.respectively.(3)combined with classic machine learning methods,an algorithm of target segmentation and recognition in railway environment is proposed.On the railroad data set,our method is proved to be eligible to perform classification task for five categories(ground,tree,linear objects,construction and track bed).The range of precision /accuracy / sensitivity are from 46.778%to 96.114%,from 92.578%to 97.823%,from59.328%to 94.715%,respectively.Compared with the current machine vision and neural network algorithm,the algorithm makes full use of prior information,with low complexity.