The Building Research Establishment (BRE) is a centre of building science in the United Kingdom, owned by a charitable organisation, the BRE Trust. BRE provides research, advice, training, testing, certification and standards for public and private sector organisations in the UK and abroad.
Augmenting Anomaly Detection with Tiny Cameras
Explore the Role of Tiny Cameras Towards Augmenting Anomaly Detection within Built Environments
Modern smart built environments are Cyber-Physical Systems (CPS) in nature. CPSs are composed of physical systems (hardware), software systems and potentially other systems (e.g., human systems). In the cyber world, anomalies are detected through analysing network packets. However, the cyber-physical world requires a different approach to monitor both network and physical worlds. An anomaly is an observation that does not conform to a normal pattern. Anomalies within built environments include intrusion, fire, variation in power consumption, unusual activation of smart devices, abnormal living patterns and so on. Traditional physical anomaly detection systems (e.g. temperature sensor monitoring afire through temperature variations) use simple sensors (temperature, humidity, vibration, motion). For example, an open window has been detected using a temperature sensor. However, as the complexity of the anomalies increases, the achieved results become less accurate. In addition, traditional sensors can be affected by noises produced by the surrounding environment. Another limitation of traditional sensors is that they can only detect measurable properties, and simple sensors cannot detect some parameters. Cameras are an advanced type of sensor that has been used mainly in surveillance tasks. Historically, in anomaly detection, the utilisation of camera sensors is limited due to multiple factors such as increased costs, comparatively larger, and privacy issues. However, tiny cameras are becoming cheaper and less than 1 inch in length.
This project investigates how to augment sensor-based anomaly detection systems with tiny cameras in a privacy-aware manner. For example, to reduce privacy invasion, camera sensors will only be activated to observe a scene if another sensor (e.g. temperature, motion) produces an abnormal result. Further, we believe tiny cameras can be used to train other senors over time to improve their anomaly detection capabilities and reduce the involvement of tiny cameras in decision-making, therefore reducing privacy concerns. This project use pre-trained object detection and computer vision models to detect anomalies and correlate them with other sensor data to improve the overall performance of the anomaly detection system. The project has the following main objectives:
- Conduct a literature review on camera systems to explore the role of the camera as a sensor in the context of anomaly detection in built environments.
- Investigate how integrating sensor-based anomaly detection with low-cost cameras can affect the overall performance.
- Identify the capabilities and limitations of the tiny camera sensor and the deployment challenges and investigate how the tiny camera can be used to (re)train other sensors over time and enhance their performance.
Team
Partners
Building Research Establishment (BRE)
Outcomes
Conference
Norah Albazzai, Omer Rana, and Charith Perera, Camera-Assisted Training of Non-Vision Sensors for Anomaly Detection, In Proceedings of the 2023 IEEE/ACM Seventh International Conference on Internet-of-Things Design and Implementation (IoTDI) 2023, pp. 452–453
