In-Vehicle Wireless Gesture Recognition System-Based Attention-Enhanced Lightweight Model

With the rapid development and wide utilization of intelligent connected vehicles (ICVs), the interaction between in-vehicle infotainment system (IVIS) and drivers has become especially crucial. In this paper, we develop the in-vehicle wireless gesture recognition (GR) system based channel state information (CSI) acquirement between IVIS and the driver’s smartphone. Limited by the smartphone’s battery and computational capacities, decreasing the CSI dimension for GR and making the model lightweight are our main objectives. Thus, we propose the subcarrier grouping selection strategy (SGSS) to reduce the dimension of CSI input from 256 to 13 through grouping the subcarriers into two sets and accomplishing CSI reshape among three selection strategies. Meanwhile, the attention-enhanced lightweight (AELite) model is designed by integrating the bidirectional gated recurrent unit (BiGRU) network with the channel attention module (CAM) for driver GR with only 8.5 million parameters. Through the experiments conducted in a real vehicle environment, the AELite model achieves the average accuracy of 94.89% and the F1 score of 94.88% both on large-scale and small-scale gestures, demonstrating exceptional recognition performance while maintaining a low number of parameters.