Learning-based Anomaly Detection for Aerial Surveillance

Ilker Bozcan

    Research output: Book/anthology/dissertation/reportPh.D. thesis


    Anomaly detection is crucial for autonomous surveillance systems. In order to increase the autonomy level of surveillance, such a system should be able to detect and alert human operators in case of the presence of unusual observations (i.e., anomalies or outliers). However, it is not a straightforward task to detect anomalies due to several challenges caused by the nature of anomalies. First, anomalies are rare compared to the majority of observations. Second, the definition of anomalies can keep change over time. Lastly, an observation can be judged abnormal or normal depending on the observation context. Therefore, a proper anomaly detection method should address these challenges, which are also valid for visual surveillance.

    This thesis formulates an anomaly detection framework for visual surveillance in a modular approach. Unlike similar studies with an end-to-end approach, the proposed framework has a modular structure that brings crucial advantages over standard anomaly detection studies. The thesis decouple outlier detection into three components: (i) perception of an environment, (ii) meta-representation of the environment, (iii) anomaly detection on meta-representations. This thesis focuses on meta-representation and anomaly detection. However, it also contributes perception of an environment as it is a preliminary step for the rest of the thesis's subject. Then, the thesis proposes a meta-representation method that indicates the spatial layout of objects in a given image. Lastly, we propose different anomaly detection methods which address different challenges in anomaly detection.

    The main application scenario is traffic surveillance, where an unmanned aerial vehicle is deployed to detect abnormalities during traffic monitoring. However, we extend use-cases with indoor surveillance and smart manufacturing to investigate different aspects of the proposed anomaly detection approach. Due to the rarity and obscurity of anomalies, the proposed approach follows the primary paradigm of anomaly detection: learning normal patterns in a given data instead of learning anomalies itself. Then, anomalous samples can be identified as they do not comfort well with the normal data. Although existing studies have an end-to-end approach where they get raw visual data as input and give anomalies as output, we propose a modular approach that is a better fit for visual surveillance than traditional methods. Moreover, our method can detect contextual anomalies which are considered normal or abnormal depending on the observation settings.
    Original languageEnglish
    PublisherAarhus Universitet
    Number of pages146
    Publication statusPublished - Dec 2021


    • Anomaly detection
    • Surveillance
    • Drone
    • Machine learning
    • Artificial intelligence


    Dive into the research topics of 'Learning-based Anomaly Detection for Aerial Surveillance'. Together they form a unique fingerprint.

    Cite this