Multi-Scale Hierarchical Conditional Random Field for Railway Electrification Scene Classification Using Mobile Laser Scanning Data

With the recent rapid development of high-speed railway in many countries, precise inspection for railway electrification systems has become more significant to ensure safe railway operation. However, this time-consuming manual inspection is not satisfactory for the high-demanding inspection task, thus a safe, fast and automatic inspection method is required. With LiDAR (Light Detection and Ranging) data becoming more available, the accurate railway electrification scene understanding using LiDAR data becomes feasible towards automatic 3D precise inspection. This thesis presents a supervised learning method to classify railway electrification objects from Mobile Laser Scanning (MLS) data. First, a multi-range Conditional Random Field (CRF), which characterizes not only labeling homogeneity at a short range, but also the layout compatibility between different objects at a middle range in the probabilistic graphical model is implemented and tested. Then, this multi-range CRF model will be extended and improved into a hierarchical CRF model to consider multi-scale layout compatibility at full range. The proposed method is evaluated on a dataset collected in Korea with complex railway electrification systems environment. The experiment shows the effectiveness of proposed model.