Privacy-aware smart camera for abnormal event detection in home environments

Abstract

Ensuring both effective monitoring and user privacy is essential in home surveillance applications. This study proposes a privacy-aware smart camera system integrating a servo-controlled mechanical housing with a dual-branch deep learning framework. That is a smart camera housing equipped with servo-controlled lids and dual operation modes, local button control and WiFi-based remote control, providing convenient usage while preventing unintended image capture. To detect hazardous household events, we constructed the EPUabInhouse dataset and proposed a dual-branch framework that integrates YOLOv8 for spatial analysis with RAFT optical flow for motion representation. Experimental results show that incorporating RAFT leads to a relative improvement of 2.02% to 4.15% in F1-score across different classes and significantly reduces background misclassification. These enhancements demonstrate the effectiveness and practical applicability of the proposed privacy-aware home surveillance system.