A Modeling Approach for Human Gesture Channels Based on Deep-Learning

Human gesture recognition has emerged as a promising application of Integrated Sensing and Communication (ISAC) in 6G networks, where gesture-induced body movements significantly reshape wireless propagation environments. However, existing statistical and deterministic channel models fail to capture fine-grained gesture dynamics and lack interpretability. In this paper, we propose a deep-learning-based framework for gesture-driven channel modeling. A Poisson neural network first predicts the number of multipath components (MPCs) associated with different body parts, followed by a Conditional Variational Autoencoder (C-VAE) that generates scattering points with spatial and temporal consistency. Based on real measurements at 28 GHz using a context-aware channel sounder with synchronized RF, camera, and Lidar data, the proposed framework reconstructs channel impulse responses (CIRs) with high fidelity. Experimental results demonstrate that the generated scattering points closely match measured distributions, while the reconstructed CIRs accurately reproduce power delay profiles.