Range-Aware Hand Gesture Recognition Using FMCW Radar and Deep Learning
Yosuke Iida, Mingyang Fan, Walid Brahim, Jianhua Ma, Muxin Ma, Alex Qi · 2023
This paper introduces a hand gesture classification implementation that combines FMCW radar and deep learning. Unlike previous work, we implement a Range-Aware methodology to automatically locate the hand range bin enabling more accurate classification. The inspiration behind our proposal is the observations made about the impact of the subject chest in the Range-Time (RT) information and the signature of the hand movement in the Range-Doppler (RD) space. Subsequently, we propose three distinct methods for hand range bin localization. Method I and II take advantage of the first observation by locating the chest and using it as a reference to select the hand bin. Method III, however, exploits the second observation about the Doppler signature of the hand movement to directly select the hand bin. The Doppler matrix resulting from each method is then fed to a CNN-based model for the hand gesture classification task. We perform subject-dependent and independent evaluations to classify six hand gestures and investigate the impact of several parameters including the type of input data, the use of an LSTM layer, and an increased range of up to 90 cm. The evaluation results show good performance even for subject-independent and from distance up to 90 cm achieving an average accuracy of 99.86% in subject-dependent evaluations and 94.23% in subject-independent scenarios.