Lightweight micro-motion gesture recognition based on MIMO millimeter wave radar using Bidirectional-GRU network

Non-contact gesture recognition is a novel form of human–computer interaction. It has broad prospects in many applications, such as Augmented Reality/Virtual Reality, smart homes and intelligent medical systems. Therefore, it has become a research hotspot in recent years. This paper investigates a lightweight micro-motion gesture recognition method based on Multiple Input and Multiple Output millimeter wave radar. We employ TI’s MMWCAS radar, comprising four cascaded AWR1243 radar boards, to collect gesture data. During the data pre-processing stage, we extract the Range-time Map, Doppler-time Map, Azimuth-time Map and Elevation-time Map of the dynamic gestures to characterize the dynamic motion features. These maps are then simplified into a one-dimensional vector to reduce data volume. We propose an 8HBi-GRU model, which combines the Bidirectional Gate Recurrent Unit (Bi-GRU) with a multi-head self-attention mechanism, to identify twelve types of micro-motion gestures using feature vectors. The model achieves an accuracy of 98.24 % , with precision and recall rates exceeding 0.97 and 0.98, respectively, for ten of the gesture types. Experimental results demonstrate that the proposed 8HBi-GRU model achieves lightweight gesture recognition rapidly and requires minimal storage space compared to image-based deep learning methods.