A Deep Learning based Hand Gesture Recognition on a Low-power Microcontroller using IMU Sensors

In this paper, we demonstrate an inertial measurement unit (IMU) based hand gesture recognition (HGR) on a low-power microcontroller (STM32L476JGY). The focus of this work is to build a reliable hardware prototype by using deep neural networks (DNN) deployed on a resource limited device. To train the DNNs, a dataset was recorded which contains accelerometer and gyroscope readings from three IMUs mounted on the fingertips. With this dataset, various neural networks (NN) were trained and analyzed. The best NN, in terms of accuracy, memory usage and latency, was then selected and ported to the microcontroller. Finally, a runtime analysis of the model has been performed on the controller. The analysis showed that a LSTM is best suited for the detection of hand gestures. The selected model achieves an accuracy of 93% and only takes up around 40 KiB of memory. In addition, the model has a throughput time of only 3.52 ms, which means that the prototype can be used in real time.