Pixel-based pipeline hardware architecture for high-performance Haar-like feature extraction

Abstract

Feature extraction, which is one of the basic tasks for pattern recognition, has often high computational cost and large memory usage. In this work, we propose a pixel-based pipeline hardware architecture for Haar-like feature extraction, implemented in 0.18 μm CMOS technology with 1.76 mm2 core area. Pixel-input speed relies on the working frequency of the image sensor so that features are extracted in real time without on-chip image buffer and complex computational procedures. The fabricated chip consumes 4.78 mW power at 1.8 V supply voltage and 12.5 MHz frequency during 30 fps VGA video input. Furthermore, a processing time of 3.07 ms per VGA frame with power dissipation of 36.25 mW at 100 MHz frequency is possible.