Anomaly Detection on the Rail Lines Using Semantic Segmentation and Self-supervised Learning

Kurzfassung

This paper introduces a novel application of anomaly detection on the rail lines using deep learning methods on camera data. We propose a two-fold approach for identifying irregularities like coal, dirt, and obstacles on the rail tracks. In the first stage, a binary semantic segmentation is performed to extract only the rails from the background. In the second stage, we deploy our proposed autoencoder utilizing the self-supervised learning techniques to address the unavailability of labelled anomalies. The extracted rails from stage one are divided into multiple patches and are fed to the autoencoder, which is trained to reconstruct the non-anomalous data only. Hence, during the inference, the regeneration of images with any abnormalities produces a larger reconstruction error. Applying a predefined threshold to the reconstruction errors can detect an anomaly on a rail track. Stage one, rail extracting network achieves a high value of 52:78% mean Intersection over Union (mIoU). The second stage autoencoder network converges well on the training data. Finally, we evaluate our two-fold approach on real scenario test images, no false positives or false negative