Synthetic Training Data Generator for Hand Gesture Recognition Based on FMCW RADAR

Hand gesture recognition has attracted a lot of attention recently. However, gesture recognition based on machine learning (ML) requires a huge training data set in order to achieve high recognition accuracy. Creating this training data set requires a significant effort. In order to solve this problem, a synthetic frequency modulated continuous waves (FMCW) RADAR data generator for hand gestures is proposed. This generator can produce a large amount of data that can be used for training the ML model, without collecting real data involving multiple people. For evaluation, 3600 synthetic samples for six hand gestures are generated with rich variations in hand size, speed and position. Those synthetic data are utilized to train the ML models for hand gesture recognition. Further, the models are tested by real data set acquired by an AWR 1642 RADAR. The results indicate that a convolutional neural network with 19 layers (VGG19) pre-trained model in conjunction with the XGBoost classifier can achieve an average accuracy of 87.53% on the test set. If only 2% of the real data is used for training, the XGBoost classifier alone achieves an average accuracy of 56.31%. But if the synthetic data set and 2% of the real data set are combined, the average recognition accuracy of the XGBoost classifier on the test data set is increased to 94.63%.