Poster Abstract: Camera-Assisted Training of Non-Vision Sensors for Anomaly Detection

Cameras are becoming pervasive and used for image classification and object detection in various applications, including anomaly detection. However, cameras pose a privacy threat and require significant power resources. To address these issues, researchers have explored non-vision sensors, but pre-training them for anomaly detection is challenging because anomalies are difficult to define and vary significantly across indoor environments. Thus, we propose a new approach to training non-vision sensors using a tiny camera and a pre-trained MobileNetV2 model. Data from non-vision sensors are labelled based on the image classification from the tiny camera, and an anomaly detection model is trained using these labelled data. The Random Forest model is used as the final model, achieving an accuracy of 95.58%.