Crowdsourced Indoor Wi-Fi REMs: Does the Spatial Interpolation Method Matter?

Crowdsourcing is a promising approach to collect the received signal strength (RSS) measurements required to construct Wi-Fi radio environment maps (REMs) via spatial interpolation methods. However, crowdsourced RSS measurements are often unreliable and especially in indoor propagation environments they are strongly affected by shadowing and fading. It is thus important to understand which spatial interpolation method is better suited for such REMs. In this paper we present an extensive empirical performance evaluation of four major spatial interpolation methods for crowdsourced indoor Wi-Fi REMs: Inverse Distance Weighting (IDW), Gradient Plus Inverse Distance Squared (GIDS), Ordinary Kriging (OrK), and Universal Kriging (UnK). We evaluate them both quantitatively by the estimation error, and qualitatively by visually observing the resulting REMs from measured Wi-Fi beacon RSSs. In our experiments we used sixteen Raspberry Pi boards and we considered different measurement location densities and spatial distributions in an office environment with four rooms. Both the quantitative and qualitative analysis of our experimental results show that there is only a negligible difference in the accuracy of the REMs constructed using the four considered spatial interpolation methods, relative to the inherent variability in the RSS reported by different devices. This suggests that, in practice, for crowdsourced indoor Wi-Fi REMs, simpler methods like IDW may be preferred over more computationally complex methods like Kriging. Importantly, our analysis also shows that, for moderate measurement location densities, the global RSS estimation error over a large indoor area with multiple rooms does not comprehensively characterize the performance of spatial interpolation methods. In addition to considering this metric, a quantitative analysis of the local error specific to each room, and a qualitative evaluation of the observed topology of the resulting REM need to be performed. This is a novel observation and suggests that more sophisticated metrics should be investigated for different engineering applications of REMs.