DCSNN: An Efficient and High-speed sEMG-based Transient-state Micro-gesture Recognition Method on Wearable Devices

Micro-gesture recognition using wearable devices is an important research topic in human-computer interaction. Surface electromyography (sEMG) is widely researched for gesture recognition due to its ability to capture muscle signals that precede actual gestures. Most existing methods are based on artificial neural networks (ANN), which may lead to high latency, high power consumption, and high memory usage when deployed on wearable devices. We propose a deep compressed spiking neural network (DCSNN) to address the challenges. The DCSNN can significantly reduce the inference power consumption and memory usage while improving recognition accuracy. In addition, we designed a linear method of action detection called leaky integrate-and-fire for transient-state action detection (TAD-LIF), which can improve the robustness of recognition systems effectively. To evaluate our method, we developed two lightweight sEMG wristbands respectively for two interaction modes, and collected two datasets from about 40 subjects. The experiment results show that DCSNN had a higher recognition accuracy than existing methods with values of 88.55% and 95.76% on the two datasets. In addition, its inference latency, power consumption, and memory usage are only about 0.4%, 0.05%, and 2% of those of popular convolutional neural network (CNN) methods. Our method enables precise, high-speed, and low-power micro-gesture recognition on a plethora of resource-constrained consumer-level intelligent wearable devices.