24GHz FMCW Radar Based Lightweight Real-time Hand Gesture Recognition System

In this paper, we propose a lightweight real-time gesture recognition system based on hand trajectory caught by 24GHz FMCW radar. To catch a reliable hand trajectory, we employ a computationally efficient local interpolation method to achieve super-precision positioning of the hand. Combined with lightweight neural networks for gesture capture and classification, high accuracy is achieved with minimal computational and resource consumption. After data augmentation for training, validation accuracy for 10 gestures can reach 97.5%, with the entire parameter size of only 0.02MB. The entire process from receiving raw radar data to outputting gesture recognition results takes only 6ms on a low-cost Microcontroller Unit.