Explore the Role of Cameras Towards Augmenting Anomaly Detection Within the Built Environment

Abstract

This thesis investigates how cameras can augment anomaly detection in built environments, focusing on their role in improving detection accuracy, reducing manual labelling requirements, and supporting system adaptability. The research adopts a multimodal approach, combining camera data with environmental sensor readings to explore how visual information can enhance non-vision anomaly detection systems. A comprehensive survey of camera-based anomaly detection establishes the growing presence of cameras in everyday environments and highlights their strengths and limitations as sensing devices. Building on this foundation, the thesis introduces a camera-assisted training framework. Short periods of visual observation generate supervisory labels for non-vision sensors, enabling accurate anomaly detection models without manual annotation. The work further advances the field through the development of SenseLess, a minimalvision system that relies primarily on non-vision sensors while invoking cameras selectively only when additional semantic context is required. During the training phase, anomaly cues and confidence estimates derived from non-vision sensors are temporally aligned with image data and combined with self-supervised visual clustering to automatically generate supervisory labels for unlabelled images. These refined labels are then used to train a vision-based anomaly detection model without manual annotation. During deployment, the system follows a sensor-first design in which non-vision sensors operate continuously and visual inference is triggered only when sensor-based predictions are uncertain. This design demonstrates how cameras can be used not as continuous monitoring devices but as targeted tools that strengthen non-vision sensing while preserving privacy. By integrating camera-assisted training, selective visual activation, and robust sensor alignment, this thesis demonstrates the potential of leveraging cameras to enhance anomaly detection within built environments. Recommendations include employing cameras strategically during calibration, using non-vision cues to govern visual activation, and adopting adaptive synchronisation techniques to maintain performance over time. Future work may extend these methods to larger and more diverse datasets, incorporate additional sensing modalities, and further investigate privacy-aware mechanisms that promote user control and acceptance.