Versatile Dataset Generation System for Hand Gesture Recognition Utilizing FMCW-MIMO Radar

We have developed a versatile dataset generation system for hand gesture (HG) recognition using frequency-modulated continuous-wave (FMCW)-multi-input-multioutput (MIMO) radar to improve the classification performance compared to conventional methods such as open dataset, other data generators using a generative adversarial network (GAN), and motion capture tools. The proposed system consists of an HG trajectory generator, an intermediate frequency (IF) signal generator corresponding to antenna locations, and a sampling timing generator without any open datasets or any motion capture data utilizing other sensors. After the training is performed by the generated dataset, the testing is carried out by actual data collected from FMCW-MIMO radar. Our findings show that the accuracy of 98% can be achieved with the generated dataset, and the proposed system is available for pretraining without using an actual dataset. Furthermore, when the mixed dataset is used for the training process, the accuracy improves by almost 37 points compared to when using the actual dataset only.