Unsupervised intrusion detection for rail transit based on anomaly segmentation

Detecting intrusions in rail transit can be challenging using traditional supervised methods, as they only detect target categories present in the training dataset and require extensive manual annotations. This paper proposes an unsupervised method for railroad intrusion detection based on anomaly segmentation, called heterogeneous uninformed students network (HUS-Net). No obstacle data is needed for training with this method, and it does not restrict identified objects to specific categories. HUS-Net utilizes a pre-trained descriptive model as the teacher network and distils its knowledge into two heterogeneous students via multi-level feature pyramid matching and reconstruction techniques. The representation discrepancy between the students and the teacher is utilized to identify railroad intrusion events and locate anomalous objects. The model is evaluated on images captured by an onboard vision system in real rail transit operating environments. Experimental results demonstrate that HUS-Net can accurately and efficiently detect intrusion events on railroads and segment invading objects, achieving better performance than other anomaly segmentation methods.