A Scalable and Efficient Hardware Accelerator for Realtime EMG-based Gesture Recognition using Deep Learning

Abstract

Real-time hand gesture recognition from surface electromyography (sEMG) is a key enabling technology for myoelectric prostheses. Deploying accurate deep learning models on resource-constrained wearable hardware remains a central challenge. This paper evaluates the OpenEye FPGA accelerator for inference of a multi-scale 1D-CNN on the Ninapro DB2 benchmark dataset recorded from healthy subjects. We characterize classification accuracy across a hyperparameter grid (kernel configuration, window size, pool size) and assess the impact of post-training INT8 quantization. OpenEye achieves inference latencies as low as 5.3 ms, well below the 125 ms perceptual threshold, with near-linear throughput scaling across PE configurations and a quantization accuracy penalty of less than 0.1%.