| dc.description.abstract | Wi-Fi sensing enables innovative applications in
healthcare and surveillance by providing continuous, contactless
monitoring through existing infrastructure. Moreover, exploiting
information from different receivers within an environment (i.e.,
different views) and using it as input to Deep Learning (DL)
models facilitates sophisticated use cases. Previous works have
demonstrated that this approach, also known as collaborative
sensing, increases sensing coverage and significantly boosts
accuracy and robustness. However, existing DL models for
collaborative Wi-Fi sensing exhibit two critical limitations: (1)
They do not consider the optimal fusion point for data from
different receivers; (2) They assign equal importance to all
receivers, irrespective of their sensing capacity. In this paper, we
address the above gaps by proposing MULTI-FI, a novel model
enhancement framework consisting of two modules. The first one
benchmarks the appropriate fusion strategy, given the charac
teristics of the sensing scenario. By leveraging physical context
information like location and orientation and the appropriate
fusion strategy, the second module instruments a model re-design
strategy. Our validation of MULTI-FI spans across several SoTA
DL models, three real-world datasets and two applications, and
shows improvements over the original model version in every
case, with average accuracy gains of 8.2% and up to 29.6%. | es |
