关键词: 手势识别, 神经网络, CMOS图像传感器, 读出电路

Abstract: Hand gesture recognition technology has garnered significant attention in the fields of human-computer interaction and virtual reality. However, vision-based gesture recognition systems is still limited by chip area and power consumption. A readout circuit suitable for the sliding convolution operation of a 3×3 convolution kernel is proposed. It includes convolution and ReLU-MaxPool dual-function circuit. The convolution circuit can achieve configurable positive and negative weights, utilizing current weighting to perform high-precision convolution operations in the current domain. The ReLU-MaxPool dual-function circuit can simultaneously achieve ReLU and MaxPool functions. It employs switch-capacitor, registers, and customized logic circuits, and combines timing design to greatly improve work efficiency and operating speed. The proposed circuit is implemented based in 55 nm CMOS process with a supply voltage of 2.5 V. The output error of the convolution circuit under different process corners is 0.25%. After Monte Carlo simulation, the mean percentage error of the convolution circuit is 0.05%, with a variance of 1.26%. By modeling the circuit errors and incorporating them into the gesture recognition algorithm, the accuracy on the test set decreased from the ideal algorithm model's 91.66% to 90.62%, with only a 1.04% impact on the results.

Key words: hand gesture recognition, convolutional neural network (CNN), CMOS image sensor, readout circuit