Readspeaker menu

Master's Thesis about Anomaly Detection Published in High-Ranking Journal

Detecting Regions of Maximal Divergence for Spatio-Temporal Anomaly Detection

What do hurricanes, heart arrhythmia, credit card fraud, and plagiarism have in common? All these events are anomalies. They do not conform to the usual patterns and rules underlying the majority of the data. Such anomalies can often disturb pattern recognition methods because they distract from the regularities in the data. On the other hand, analyzing the anomalies can lead to a better understanding of the data and the processes being observed. Regarding climate research, for instance, extreme and uncommon weather events are of particular interest for gaining new insights about global climatic relationships and for monitoring the health of our planet. Moreover, there are many application scenarios for which only the unusual events are relevant: anomalies in an electrocardiogram, suspicious credit card transactions, or stylistic anomalies in written texts, which might be indicative of plagiarism.

Maximally Divergent Intervals for Anomaly DetectionA team of researchers from the Computer Vision Group at the Friedrich Schiller University Jena has now developed a method for detecting such anomalies in various types of data. The algorithm called "Maximally Divergent Intervals (MDI)" achieves this by searching for regions in time-series data whose statistical distribution differs significantly from the rest of the data. This method is unsupervised and does not need to be trained on nominal data in advance. Moreover, it can handle data that does not only possess a temporal but also a spatial dimension as well as simultaneous measurements of multiple different variables, e.g., wave height, wind speed, and air pressure.

The researchers successfully applied this method for detecting the major North Sea storms from the past 50 years in half a million of metocean measurements as well as unusual activities in surveillance videos and stylistically anomalous paragraphs in books. The theory underlying the new method and the results from this work have recently been published in the high-ranking pattern recognition journal IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI).

Based on a master's thesis

The idea behind the MDI algorithm has originally been developed in the group BACI (Biosphere-Atmosphere Change Index) by the postdoc researchers Dr. Erik Rodner and Dr. Yanira Guanche Garcia from the Computer Vision Group Jena directed by Prof. Dr. Joachim Denzler. This research was part of the group's work on the detection of extreme weather events in climate data. The actual research and development leading to the final algorithm was then done by Björn Barz, at that time a master student and now a Ph.D. student at the Computer Vision Group. In his master's thesis he developed the essential theory and components without which the MDI algorithm could not be applied efficiently and effectively in real application scenarios. It is a rare exception that a master's thesis results in a publication in such a high-ranking journal as TPAMI.


B. Barz, E. Rodner, Y. Guanche Garcia, J. Denzler.
"Detecting Regions of Maximal Divergence for Spatio-Temporal Anomaly Detection."
IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018.
DOI: 10.1109/TPAMI.2018.2823766
Available online as Open Access publication

previous | next