Tiny TCN model for Gesture Recognition With Multi-point Low power ToF-Sensors

Gesture recognition is currently an important field of research, as it provides the basis for the next generation of Human Machine Interface. Especially recognizing gestures with touch-less technology has several advantages including, for example, the possibility of working with gloves either outside or in sterile environments. In this work, we present a tiny machine learning neural network model exploiting a novel, miniaturized and low power Time of flight (ToF)-sensor from STMicroelectronics. The algorithm is based on a combination of temporal convolutional networks and is optimized to work on an embedded small and low power ARM cortex-M4 processor with a few kilo Byte of RAM, as employed in every recent Bluetooth Low Energy module. A dataset has been collected with the ToF sensor for training and evaluation of the proposed algorithm. The model has been evaluated for accuracy and a trade-off between computational resources and accuracy is presented. The gesture recognition achieves a very high accuracy of over 96% over 6 different gestures.