A 213.7-µW Gesture Sensing System-On-Chip With Self-Adaptive Motion Detection and Noise-Tolerant Outermost-Edge-Based Feature Extraction in 65 nm

Abstract

This letter presents a low-power motion gesture recognition system-on-chip (SoC) for smart devices. The SoC incorporates a low-power image sensor and a memory-efficient outermost-edge-based gesture sensing DSP. The DSP utilizes a self-adaptive motion detector that automatically updates a motion-pixel threshold for accurately sensing hand movements. A convolution-based noise-tolerant feature extraction (FE) technique is also developed for preventing detection errors caused by random noises in the images from the low-power sensor. The FE architecture is highly accelerated utilizing parallelisms and pipelining for achieving low-latency real-time gesture recognition. Measurements from a test chip fabricated in 65-nm CMOS show that the SoC consumes 213.7 µW with only 3-µW dynamic power at 30 f/s. The SoC occupies only 0.54 mm2, making it well applicable for wearable devices and sensor nodes. The image sensor is fully operational down to 0.6 V while the DSP can be scaled down to 0.46 V. The average recognition accuracy of the system is 85% while the latency is 1.056 ms.