A Low-Power Vision System With Adaptive Background Subtraction and Image Segmentation for Unusual Event Detection

This paper presents a smart ultra-low power vision system targeted to video surveillance applications. The sensor embeds a low-level image processing technique that autonomously detects unusual events occurring in the scene, relying on adaptive background subtraction. The resulting binary image is then directly segmented by an FPGA, which triggers the higher layer of processing, transferring only aggregate feature information. The on-board processing relieves the rest of the vision system from expensive computation. The $104\times 104$ pixels vision chip consumes $80~\mu \text{W}$ at 30 frames/s, while segmentation dramatically cuts down the amount of data to be transferred, resulting in an extremely low-power system suitable for embedded applications.