Human Gesture Recognition Based on CT-A Hybrid Deep Learning Model in Wi-Fi Environment

Human gesture recognition has become an important aspect of human–computer interaction due to the rapid development of human behavior sensing technology in Wi-Fi environments. Although Wi-Fi-based gesture recognition systems have achieved good accuracy within specific domains, they still have limitations in terms of cross-domain capability. In light of this, this article aims to explore methods that can achieve high recognition accuracy within specific scenes while also maintaining cross-scene capability. To address this challenge, we propose a hybrid deep learning model that leverages a combination of convolutional neural network (CNN) and the encoder module in the Transformer. This model takes into consideration the spatial localization characteristics and long-distance dependence of gestures, which improves its ability to model the spatiotemporal features in the body-coordinate velocity profile (BVP) series. In addition, we enhance the model’s modeling effect on spatiotemporal features in BVP series by extracting low-dimensional vectors containing a significant amount of classification information. These vectors are then fed into the Adaboost module for ensemble learning. Finally, a strong classifier is used to compute the class of gestures. To evaluate the performance of our proposed model, we conduct experiments on a common dataset. The results demonstrate that our model achieves an average accuracy of 96.78% and 88.27% in in-domain and cross-domain cases, respectively. This indicates the superiority and effectiveness of the proposed approach.