Gesture Recognition Using MIMO Radar Point Clouds With Targeted Signal Processing for Resource-Limited Platforms

Utilizing gestures for human–machine interaction offers numerous benefits for both products and users. Hand gesture recognition (HGR) systems, particularly those using radar sensors, are gaining attention due to their potential for real-time applications. However, mainstream methods often rely on unrefined 2-D or 3-D radar signals, which are typically large and redundant, requiring the use of large neural network classifiers. This high computational demand limits their deployment on low-cost, resource-constrained edge platforms. In this article, we propose a low-cost, close-range HGR system based on a 24-GHz frequency-modulated continuous-wave (FMCW) multi-input–multi-output (MIMO) radar designed for resource-constrained environments. By applying beamforming, constant false alarm rate (CFAR) target detection, and feature extraction, our experiments show that targeted signal processing on dimensions, such as range, velocity, and angle, can effectively classify gestures. This approach achieves a recognition accuracy of 99.60% for four gesture types using a lightweight convolutional neural network (CNN). By focusing on efficient radar signal processing rather than complex neural networks, we successfully deployed the system on an embedded platform. The system achieved accuracy rates of 99.36%, 95.00%, and 95.64% for a well-trained user and two nontrained users, respectively.