Improved Multi-Output AI for Indoor Localization Application Based on Real Data Measurements

Indoor localization applications are increasingly used for tracking targets, particularly mobile ones, but tracking in non-line-of-sight (NLOS) conditions presents significant challenges, especially when relying on Received Signal Strength Indicator (RSSI) measurements. The variability of RSSI values is notably higher in NLOS environments. This study investigates multi-output AI system integration for real indoor localization implementation, focusing on optimizing model performance during the offline phase and assessing their real-world applicability. The Artificial Neural Network (ANN) and Random Forest (RF) models were tested in different environments, with results demonstrating their effectiveness in specific settings. The study area covers 193.25 m<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>, using a communication protocol that adjusts transmission power to track target mobility dynamically. Two phases of optimization were conducted: offline and online. Following optimization, RF was applied for tracking on the first floor, and ANN was selected for the second floor. On the second floor, ANN achieved an RMSE of 2.06 m, while RF outperformed on the first floor with an RMSE of 1.94 m. The offline phase played a crucial role in refining model accuracy, and once transitioned to real-time implementation, the optimized models demonstrated robust and efficient localization performance. These findings emphasize the importance of combining offline optimization with real-environment implementation to develop accurate and reliable indoor positioning systems.