RaFFD: Resource-aware Fast Foreground Detection in embedded smart cameras

摘要

Embedded smart cameras have made a dramatic shift towards distributed surveillance systems by combining sensing, processing and communicating on a single platform. A critical issue in embedded smart cameras is resource-limited, which poses great challenging in designing fast and efficient vision algorithms. In this paper, we explore light-weighted foreground detection in resource-limited embedded smart cameras. More specifically, we propose RaFFD (Resource-aware Fast Foreground Detection) that reduces the computation and storage overhead in foreground detection. Observing that computation and storage overhead increase proportionally to its pixel manipulation, RaFFD deals with the target's contour points instead of the whole image. RaFFD incorporates a contour-based detection with dynamic background update, ensuring accurate foreground detection and address the bottlenecks of processing speed. We have implemented RaFFD on the our embedded smart camera platform based on CITRIC architecture. Our experimental evaluation shows that RaFFD can detect foreground with close to 95% accuracy and 6% false alarm. Even in an challenging scenario with illumination and vibration influence, RaFFD can still maintain the good robustness. Compared to the recently detection method oriented to embedded systems, RaFFD can increase processing speed to approximately twice and decrease memory consumption by 68%.