A versatile recognition processor employing Haar-like feature and cascaded classifier

This paper presents a versatile recognition processor that performs detection and recognition of image, video, sound and acceleration signals, while dissipating 0.15 muW/fps to 0.47 mW/fps. Given the low power dissipation of sub-mW/fps, this processor is suitable for use in portable electronics and wireless sensor networks (WSN). For instance, it detects human faces from a QVGA image with 81% accuracy and consumes 0.47 mW/fps. Power consumption is 57times lower than that of conventional object recognition processors with comparable accuracy. A fair comparison, by taking technology differences into account, shows greater than 8times power efficiency. This processor detects speech from very short and low quality sound signals (72 ms in 10 s, 8 kHz, 8 b) recorded by a microphone in a sensor node. It also recognizes human activities such as walking, reading and typing from short and low quality 3D acceleration signals (2 s in 10 s, 50 Hz, 8 b) taken by an accelerometer. Recognition accuracy is over 90% in both applications. The versatility and low-power dissipation are attributed to optimal VLSI design from algorithm to architecture and circuit levels.