Extended kernel-based location fingerprinting in wireless sensor networks

Fingerprinting localization is to estimate a mobile terminal's location using its online received signal strength (RSS) measurement and offline RSS database originated from multiple access points (APs). Kernel-based fingerprinting localization is such a competitive algorithm. However, all training data need to be considered in its offline model learning stage. This render high risks for overfitting. To alleviate this, we suggest to apply clustering to the localization region of interest first and then use kernal-based fingerprinting localization for each cluster. A byproduct of clustering is that the computational load for each cluster is also significantly reduced. To further reduce the computational load within each cluster, we also suggest to apply principal component compression to the raw RSS measurements to reduce their dimensionality. The rationale for applying principal component compression is that the distributions of the RSS measurements at all calibration points (CPs) within each cluster will be more similar after clustering. Performance evaluation using both simulated data and real data show that the extended kernel-based fingerprinting localization using clustering and principal component compression have better location estimation accuracy and less computational load.