NFLOW-Based Arbitrary Subway Obstacle Detection Method

Obstacle detection is vital to the safe operation of the metro. Obstacle detection methods based on supervised methods have high accuracy and high real-time performance, while most methods based on traditional image processing have low accuracy and poor real-time performance. However, supervised learning-based methods can only detect partially trained obstacles. In order to detect all kinds of obstacles, we propose an unsupervised learning-based approach. The pixel-level features are extracted by a pre-trained feature extraction network, and the density distribution of the features is estimated using normalizing flows. Furthermore, we assume that features with low probability are obstacle features. Based on the data collected in the real operating environment of the metro, our unsupervised method achieved 98.8963% DET-AUROC, 98.94% DET-precision, 99.99+% DET-recall, 99.7557% SEG-AUROC, and 98 fps inference speed. Additionally, our method only requires obstacle-free images for training purposes and can detect any kind of obstacle.