Accurate Passive Radar via an Uncertainty-Aware Fusion of Wi-Fi Sensing Data

TL;DR

The paper tackles accurate passive radar using Wi-Fi sensing by marrying variational auto-encoding to uncover a latent distribution $p(m{Z}|m{X})$ that generates CSI observations with an evidential deep-learning classifier to quantify uncertainty and detect unseen activities. It investigates multiple architectures (No-Fused, Early-Fusing, Delayed-Fusing, and VAE-based variants) to fuse multi-antenna CSI data, finding that Delayed-Fusing delivers the best activity recognition accuracy (~0.95) while offering robust out-of-distribution handling and a path to semantic latent-physics interpretation. Experimental results on a real indoor Wi-Fi dataset show substantial improvements over baselines and reveal meaningful latent-cluster structures that align with physical phenomena like movement, obstacles, and speed. The work advances flexible, semantically interpretable characterizations of rare or unseen events (black-swan scenarios) and sets the stage for more sophisticated uncertainty-aware and neuro-symbolic reasoning in passive radar systems.

Abstract

Wi-Fi devices can effectively be used as passive radar systems that sense what happens in the surroundings and can even discern human activity. We propose, for the first time, a principled architecture which employs Variational Auto-Encoders for estimating a latent distribution responsible for generating the data, and Evidential Deep Learning for its ability to sense out-of-distribution activities. We verify that the fused data processed by different antennas of the same Wi-Fi receiver results in increased accuracy of human activity recognition compared with the most recent benchmarks, while still being informative when facing out-of-distribution samples and enabling semantic interpretation of latent variables in terms of physical phenomena. The results of this paper are a first contribution toward the ultimate goal of providing a flexible, semantic characterisation of black-swan events, i.e., events for which we have limited to no training data.

Figures (10)

• Figure 1: Magnitude of the csi collected by one antenna while a person is walking. Magnitude values are reported in arbitrary units, as measured inside the Wi-Fi chipset.
• Figure 2: Plate notation of a kingma_AutoEncodingVariationalBayes_14, with dashed lines denoting the variational approximation. $\bm{\theta}$ are the true yet unknown parameters of the $Z$ distribution which generated the data $X$, while $\bm{\phi}$ are the learnt parameters.
• Figure 3: No-Fused-1 architecture. The output of the mlp is over three activities only for the purpose of showing that it is a Dirichlet distribution. Architectures No-Fused-2, No-Fused-3, No-Fused-4 focus on antenna 2, 3, and 4, respectively.
• Figure 4: Early-Fusing architecture. The output of the mlp is over three activities only for the purpose of showing that it is a Dirichlet distribution.
• Figure 5: Delayed-Fusing architecture. The latent space representation of the csi of every antenna is first extracted from its corresponding vae, and then fused together at the input of the mlp. The output of the mlp is over three activities only for the purpose of showing that it is a Dirichlet distribution.