Railway track anomalies require prompt identification and intervention. To detect these defects, LiDAR data are gathered, and are then validated by an expert: this process is slow and inefficient. Delayed intervention can lead to additional damage, worsening costs and safety.
A client working in the sector of infrastructures management, reached out to request the development of an anomaly classification algorithm. This algorithm had to be designed to work in conjunction with an existing threshold-based algorithm, that detects anomalies in railway tracks, but has a very high false positive rate. The new algorithm had to classify these flagged anomalies and identify true anomalies from false positives.
To tackle this challenge, I developed an algorithm that employs convolutional neural network and takes the LiDAR coordinates as input. The algorithm included strong priors regarding the defects shape and size, assessed by dataset inspection. This approach has led to an algorithm that could be efficiently trained even with a relatively small dataset. The deployed method has demonstrated accuracies in classifying anomalies which are comparable to those achieved by human experts, while remaining dramatically faster.