Non-invasive air-writing using deep neural network

This paper compares the inference performance of different deep neural networks executed on hardware with limited memory and computational resources. Performance comparison is done between densely connected networks (DNN), convolutional neural networks (CNN), and a long-short term memory network (LSTM) trained to classify hand-written characters on the air. Signals from an accelerometer and a gyroscope are sampled from a MEMS sensor when drawing the symbols. The inference is executed directly on the device equipped with an STMF401 microcontroller. The figures of merit used for the comparison are memory occupation, inference time, energy consumption, and classification accuracy.