Fishing Trawler Event Detection: An Important Step Towards Digitization of Sustainable Fishing

Abstract

Detection of anomalies within data streams is an important task that is useful for different important societal challenges such as in traffic control and fraud detection. To be able to perform anomaly detection, unsupervised analysis of data is an important key factor, especially in domains where obtaining labelled data is difficult or where the anomalies that should be detected are often changing or are not clearly definable at all. In this article, we present a complete machine learning based pipeline for real-time unsupervised anomaly detection that can handle different input data streams simultaneously. We evaluate the usefulness of the proposed method using three wellknown datasets (fall detection, crime detection, and sport event detection) and a completely new and unlabelled dataset within the domain of commercial fishing. For all datasets, our method outperforms the baselines significantly and is able to detect relevant anomalies while simultaneously having low numbers of false positives. In addition to the good detection performance, the presented system can operate in real-time and is also very flexible and easy to expand.