Discovering unknowns: Context-enhanced anomaly detection for curiosity-driven autonomous underwater exploration

•Anomaly detection for unknowns towards to autonomous underwater exploration.•Autoencoder and autoregressive network to identify anomalies in unstructured dynamic underwater moving views.•Novel context-enhanced autoregressive network to learn feature dependence.•Patch learning paradigm to build an accurate latent feature space.•Validation on two benchmarks, simulation, and real data, outperforms state-of-arts.