Abstract
Accelerometers are nowadays included in almost any portable or mobile device, including smartphones, smartwatches, wrist-bands, and even smart rings. The data collected from them is therefore an ideal candidate to tackle human motion recognition, as it can easily and unobtrusively be acquired. In this work we analyze the performance of a hand-gesture classification system implemented using LSTM neural networks on a resource-constrained microcontroller platform, which required trade-offs between network accuracy and resource utilization. Using a publicly available dataset, which includes data for 20 different hand gestures recorded from 10 subjects using a wrist-worn device with a 3-axial accelerometer, we achieved nearly 90.25% accuracy while running the model on an STM32L4-series microcontroller, with an inference time of 418 ms for 4 s sequences, corresponding to an average CPU usage of about 10% for the recognition task.