Robust Registration of Rail Profile and Complete Detection of Outliers in Complex Field Environment

The accurate measurement of rail wear is critical in track quality inspection. Usually, by matching the measured profile with the standard one based on the unworn rail waist double circle segment (DCS), and comparing their railhead differences, we obtain the rail wear. However, in the complex field environment, this task becomes tricky. For profile registration, one is that the location of rail waist becomes difficult with lots of outliers mixed in the profile, and the other is that the obtained rail waist could be polluted or incomplete. For wear measurement, the outliers scattered on the railhead could cause serious errors. This paper is devoted to solving the above problems. For problem 1, firstly, the rail waist is located after a preprocessing procedure. Then, utilizing the proposed hybrid model called R-H-ICP, we realize the profile registration with a process from coarse to fine. For problem 2, by using the standard profile and the reconstructed one with wear constraints to serve as the template separately, and computing the distance from each point of railhead on measured profile to the template, both the distinct and the inapparent outliers are detected precisely. The efficiency and superiority of proposed methods were verified by vast experiments. For the former, compared with DCS and ICP, the R-H-ICP not only improves the profile utilization ratio obviously, also guarantees the registration accuracy as far as possible. For the latter, the F1-Measure average score of outlier detection reaches 0.998, which outperforms some classical models markedly.

Automated summary

This paper focuses on solving the challenges of rail wear measurement in complex field environments. It proposes a solution that includes preprocessing to locate the rail waist, a hybrid model called R-H-ICP for profile registration, and a method for accurately detecting outliers. Extensive experiments validate the efficiency and superiority of the proposed methods.