Abstract:

Complementary to the fine-grained channel state information (CSI) and coarse-grained received signal strength indicator (RSSI) measurements, the mid-grained spatial beam attributes (i.e., beam SNR) during the millimeter-wave (mmWave) beam training phase were recently repurposed for Wi-Fi sensing applications such as human activity recognition and indoor localization. This paper proposes a multi-band Wi-Fi sensing framework to fuse features from both CSI from 5 GHz bands and the mid-grained beam SNR at 60 GHz with feature granularity matching that pairs feature maps from the CSI and beam SNR at different granularity levels with learnable weights. To address the issue of limited labeled training data, we propose to pretrain an autoencoderbased multi-band Wi-Fi fusion network in an unsupervised fashion. For specific sensing tasks, separate sensing heads can be attached to the pretrained fusion network with fine-tuning. The proposed framework is thoroughly validated for three sensing applications using in-house experimental datasets: 1) pose recognition; 2) occupancy sensing; and 3) indoor localization. Comparison to a list of baseline methods demonstrates the effectiveness of granularity matching. An ablation study is performed as a function of the amount of labeled data, the latent space dimension, and learning rates