An RSSI Indoor Localization Technique Using Segmented CNN Regression

In wireless communication standards such as IEEE 802.11az, indoor localization techniques are expected to achieve submeter accuracy for security and/or privacy applications. Motivated by this background, we propose a new receive signal strength indicator (RSSI)-based localization technique using segmented convolution neural network (CNN) regression, which can be applicable for legacy devices. Specifically, assuming a testbed partitioned into certain subregions, the proposed technique estimates a subregion as a coarse estimate by using a K-nearest neighbors (KNN) algorithm, and it refines the position estimate by performing a CNN regression whose model is trained per each subregion. The experiment results using measurement data verified that the proposed technique achieves mean absolute error (MAE) of 0.34m by reducing more than 0.2m from those of the conventional cascade-connected artificial neural network (CC-ANN) and extreme gradient boosting (XGBoost) methods. In addition, the simulation results confirmed that the proposed technique reduces costs needed to make a fingerprint database to 7% of that of the XGBoost in indoor propagation channels generated from a ray-tracing model.