Custom Rail Classifier with Digitale Schiene

At Pointly, we are excited to present our successful collaboration with Digitale Schiene Deutschland, part of DB InfraGO AG. We have successfully completed the second iteration of training a Deep Learning Custom Classifier for assets of Deutsche Bahn. In this project, we worked together on the development and implementation of a tailored classifier specifically trained on continuous mobile mapping scans of railway tracks.

Scope of the project. In the first iteration, the project involved the iterative creation of training data in multiple batches, as well as the training and deployment of a custom classifier. A key aspect was the classification of railway tracks to extract both the rails and objects along the tracks. In the recently completed second iteration, the objective was to adjust and expand the class catalog. The final work package included retraining the classifier with new training data and optimizing pre-processing and post-processing to not only extract rails cleanly but also to distinguish between top and lower rails and correct any mixtures of vegetation extensions and overhead lines.

Iteration 1 against iteration 2

Iteration 1: the rails as one class, and the signs, signals and transformer boxes merged into another

Iteration 2: after retraining, the pole objects are told apart and the top rail is its own class

Optimized and unified extraction of top rails, in blue, based on the initial rail classification, in purple

Data provided. For training in the second iteration, homogeneous datasets from different areas in Germany were selected on which the classifier was trained. Individual datasets were excluded due to differing point densities or sensor properties. Two batches were created: the first batch served as the base, which was fully utilized in training, and a second batch with specifically selected scenes and objects, tailored to enhance accuracy on the newly introduced object classes. The results achieved were impressive. The Custom Classifier trained for Digitale Schiene Deutschland runs on the Pointly Platform, the trained model is theirs, and it achieves consistent high accuracies in classification for the target scenes. Well represented classes showed accuracies of up to 95 to 98 % in the validation dataset, while some of the newly introduced classes show further potential for improvements with additional training data and post-processing improvements.

Future direction. The latest updates of the Pointly Platform include vectorization visualization and editing capabilities and are expected to be of greater significance for further projects. In the field of vectorization, Pointly intends to work on automatically extracting vector models such as for railway tracks, poles or signs while deriving additional metadata for the objects such as type, height, width, etc.

The successful collaboration in our second project with Digitale Schiene Deutschland is a milestone for us at Pointly. We are proud to be part of this foundational project and look forward to delivering innovative solutions for our customers in the future.