Mobile mapping using train-mounted scanners has become the standard method to acquire LiDAR railway point clouds. The point clouds are used for asset management of railway objects or 3D modelling, for example. These applications requires high accuracy of the point clouds, correctly representing real-world geometries and spatialities. Since the point clouds are obtained at different times, as well as measurement errors posed by GNSS signal quality, the position of the point clouds may not fully align. Iterative Closest Point (ICP) is used as the industry standard to align the point clouds, but struggles with scenes containing repetitive or symmetrical structures. These features are very commonly found on the railways: sleepers, catenary poles, or the tracks themselves. Thus, ICP registration results are not perfect, requiring additional manual fine alignment, using local evaluating sections. This is a time consuming process that requires a lot of attention, time and costs. In order to alleviate this, two automatic procedures were proposed, and their viability were examined. 

The first method - Direct Distance Evaluation - calculates the two-way Chamfer Distance between nearest points of two point clouds. It is simple and quick to implement, using the same evaluating sections for manual adjustment as input data, with misalignment results obtained for every evaluating sections along the track. Additional properties such as choosing the best cloud for further adjustment was also included. However, the computed misalignments are usually greater by a few centimeters compared to the true manual adjustments. Furthermore, it is sensible to objects partially visible in different scans, driving up the magnitude of misalignment between the points, when in reality it should not be considered misaligned.

The second method - Geometry-based Evaluation - uses objects present in the point clouds to calculate the misalignment. Using the classified point clouds from the in-house AI classification model of GeoNext as input, the method aims to find the misalignment via estimating the shift in positioning of railway objects between different scans of the same scene. Emphasis was placed on using railway sleepers as the object of interest, due to their abundance in railway, stability, and strong geometrical form (defined edges and corners). Each sleeper is considered one cluster with its own bounding box. The horizontal misalignments are then the shift in the center point of the bounding boxes representing the same sleeper in two clouds. For each point cloud, a grid of cells was also created, and the vertical misalignment is the distance of a ray orthogonal to the source cloud, between a cell on the source cloud grid to the reference cloud grid. Results for this method are much closer to the manual adjustments. However, effects from outlying estimates, choice of clustering parameters or inaccurate formation of bounding boxes could be seen in the results. 

Both methods displayed different advantages and disadvantages, however, the Geometry-based Evaluation method is recommended to be further refined and improved upon. The method works for larger sections of track compared to the evaluating sections used for manual adjustment. This can reduce the time and costs needed to annotate sections. Additional domain knowledge and optimisation could lead to better choices for parameters used in the algorithm. Furthermore, with the increasing accuracy of the AI classification results, it is expected that the method can be further developed and achieve more trustworthy results.