Dynamic hand gesture recognition based on mm-wave radar

Abstract

Research on hand gesture recognition (HGR) technology based on Millimeter- wave (mm-wave) radar is of great significance for expanding human-computer interaction (HCI) application scenarios and building intelligent terminals. How to efficiently and accurately extract the features of mm-wave radar gesture signals and design a classification algorithm suitable for gesture recognition is the key point of gesture recognition technology research. Focusing on the above problems, this dissertation conducts research on gesture recognition algorithms based on the Frequency-Modulated Continuous Wave (FMCW) radar system.

An end-to-end gesture recognition algorithm that combines deep Convolutional Neural Networks (CNN) and long short-term memory (LSTM) network is introduced, which uses CNN for feature learning and LSTM for dynamic gesture modeling. A model based on CNN and Temporal Convolutional Network (TCN) is introduced, which uses CNN for spatial and short-time modelling, and TCN for long-term modelling, and adjusts the structure of TCN to save computation.

A series of experiments were designed and carried out for the CNN-LSTM model and the CNN-TCN model respectively, and the gesture recognition performance of the two algorithms was verified from the aspects of accuracy and calculation complexity. Experiments show that both have good accuracy, among which the CNN-LSTM network has a slightly higher accuracy, and the CNN-TCN model is more computationally efficient.