Low-Complexity Gesture Recognition based on FMCW Radar

In recent years, mmWave radar-related applications have attracted a lot of attention. Radar-based gesture recognition plays an essential role among them. This paper describes how to create datasets from radar data and presents several recognition pipelines. In addition, this paper focuses on comparing the advantages, disadvantages and differences between Principal Component Analysis (PCA) and Convolutional Neural Networks (CNN) in extracting features from datasets. The experimental results show that combining PCA with a Random Forest classifier can obtain up to 99.5% accuracy, which is comparable to CNN models. The proposed combination also has lower complexity and shorter runtime than the CNN model in the prediction phase. This would benefit real-time systems for gesture recognition.