Program designed for real-time visual anomaly detection during ERC
Input images are processed by neural network based on Wide ResNet-18 backbone.
Instead of recognising predefined classes of anomalies, the model learns what normal terrain should look like and identifies regions that differ significantly from the training distribution.
Images were collected at Kąkolewo Airport Campus of Poznań University of Technology, inside a geodome simulating Mars-like terrain. The dataset contain normal terrain images and anomalies (consisting of both various objects and human images)
- Python
- PyTorch
- Anomalib
- OpenCV
- NumPy