Two-Way Long Short-Term Memory Architecture for Non-Contact Hand Gesture Identification
Arash Shokouhmand, Negar Tavassolian · 2022 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting (AP-S/URSI) · 2022
This study proposes a novel method for the classification of hand gestures using a frequency-modulated continuous-wave (FMCW) radar. A two-way long short-term memory (TW-LSTM) architecture is introduced to categorize the movements recorded by an FMCW radar into 12 standard types of hand gestures provided by IEEE DataPort. The TW-LSTM method initially segments the physical displacement features of 4,600 movements into smaller frames. It then applies intra- and inter-frame temporal operations on the data to identify the type of movement. Experiments on four different subjects report an F1 score and an average accuracy of 84.02% and 84.0% for the classification of hand gestures respectively. It is also demonstrated that the true positive rates for all subjects fall within the range of 71.0%-91.0%, showing excellent consistency between predicted and true labels of hand gestures.