Object Trajectory Estimation with Continuous-Time Neural Dynamic Learning of Millimeter-Wave Wi-Fi

In this paper, we leverage standard-compliant beam training measurements from commercial millimeter-wave (mmWave) Wi-Fi communication devices for object localization and, specifically, continuous trajectory estimation and prediction. The main challenge is that the sampling of beam training measurements is intermittent, due to the beam scanning overhead and the uncertainty of the transmission instant caused by the contention over the wireless channel. In order to cope with this intermittency, we devise a method to assist the localization by exploiting the underlying object dynamics. The method consists of a dual-decoder neural dynamic learning framework that reconstructs Wi-Fi beam training measurements at irregular time intervals and learns the unknown latent dynamics in a continuous-time fashion powered by the use of an ordinary differential equation (ODE). Utilizing the variational autoencoder (VAE) framework, we have derived a modified evidence lower bound (ELBO) loss function for the dual-decoder architecture that balances the unsupervised waveform reconstruction and supervised coordinate estimation tasks. To evaluate the proposed method, we build an in-house testbed consisting of commercial 802.11ad routers, with a TurtleBot as a mobile user, and collect a real-world mmWave Wi-Fi beam training dataset. Our results demonstrate substantial performance improvements over a list of baseline methods, further validated through an extensive ablation study.