This brief presents a wireless smart glove based on multi-channel capacitive pressure sensors that is able to recognize 10 American Sign Language gestures at the edge. In this system, 16 capacitive sensors are fabricated on a glove to capture the hand gestures. The sensor data is captured by a 16-channel CDMA-like capacitance-to-digital converter for training/inference at the edge device. Unlike the conventional approach where the capacitive information is recovered before further signal processing, our proposed system approach takes advantage of the capability of the machine learning (ML) algorithms and directly processes the code-modulated signals without demodulation. As a result, it reduces the input data throughput fed into the ML algorithms by $20\times $ . The on-site ML implementation significantly reduces decision-making latency and lowers the required data throughput for wireless transmission by at least $4\times $ . The highest testing classification accuracy of our system achieved is 99.7%, with a <0.1% difference from the conventional demodulated sensing scheme.

中文翻译：

---

一种无线多通道电容式传感器系统，用于在边缘使用 AI 进行高效的基于手套的手势识别

本简介介绍了一种基于多通道电容式压力传感器的无线智能手套，该手套能够在边缘识别 10 种美国手语手势。在该系统中，16 个电容式传感器制作在手套上以捕捉手势。传感器数据由类似 CDMA 的 16 通道电容数字转换器捕获，用于在边缘设备上进行训练/推理。与在进一步信号处理之前恢复电容信息的传统方法不同，我们提出的系统方法利用机器学习 (ML) 算法的能力，直接处理编码调制信号而无需解调。因此，它通过以下方式减少了输入 ML 算法的输入数据吞吐量 $20\times $ . 现场 ML 实施显着减少了决策延迟，并将无线传输所需的数据吞吐量降低了至少 $4\times $ . 我们系统实现的最高测试分类精度为 99.7%，与传统解调传感方案的差异小于 0.1%。