An Improvement of Accuracy for FMCW Radar Based Gesture Recognition System Using Deep Learning Models

This paper presents a gesture recognition system using FMCW radar and deep learning models. The system uses a FMCW radar to collect motion data, and explores various deep learning models, including CONV1D (1D Convolutional), LSTM (Long Short-Term Memory), and Transformer. Two distinct datasets including a public dataset and a self-collected dataset collected by separated groups are used to train and test for recognition models. By adding a portion of the self-collected dataset with the public dataset in training the accuracy of all three models are improved significantly. Among the evaluated models, CONV1D and LSTM exhibit the highest accuracy (99.17%) but LSTM has minimal accuracy reduction during cross-dataset evaluation, making it a suitable candidate for radar gesture recognition.