Maritime Anomaly Detection - A Probabilistic Deep Learning approach using GeoTrackNet and Transformers

Nielsen, Alexander Førgaard

4. semester

Artificial Intelligence, Vision and Sound, MSc.

2025

2025-06-04

33 pages

This report contains an empirical evaluation of the unsupervised Deep learning framework GeoTrackNet, originally developed for maritime anomaly detection of vessel trajectories. In contrast to the original paper, which lacked a labeled dataset, this project introduces a labeled evaluation to quantitatively asses performance of the model based on ground truths. To provide a comparison and potential improvement to this state-of-the-art model, a transformer encoder block is introduced in order to allow long-term dependency modeling using attention, inspired by models such as the TrAISformer. When both models have been trained on normal behavior and learned to recognize what normal behavior is, then a contrario detection is applied. This method uses a threshold $\epsilon$ to determine allowed number of false alarms. The models have been evaluated under a series of these values where classification metrics have been used to determine performance. The findings of this project indicate that performance difference between the original and extended model is negligible under curent training conditions. However, limited time for hyper parameter tuning, especially important for transformer-based models, may have hindered the potential and thus the results gathered.

Download
View record in AAU Student Projects