Improving Indoor Positioning Accuracy Using a Hybrid Wi-Fi RSSI Model with KNN, Kalman Filter, and MLP

Abstract

Indoor Positioning Systems (IPS) have become increasingly important due to the limitations of GPS in indoor environments. Among the different technologies, Wi-Fi RSSI is widely used because it is low-cost and works with existing infrastructure. However, RSSI signals are unstable and influenced by noise, multipath propagation, and environmental dynamics. The RADAR algorithm, one of the earliest Wi-Fi fingerprinting techniques, provides a baseline for indoor localization by comparing real-time RSSI measurements with a prebuilt radio map. While RADAR is simple and computationally efficient, it suffers from limited adaptability and accuracy in dynamic environments. In our experiments, RADAR achieved a mean positioning error of about 1.4 meters, showing its vulnerability to signal fluctuations. To improve stability, we explored a hybrid approach that combines KNN, the Kalman Filter, and a Multi-Layer Perceptron (MLP). KNN provided a simple fingerprinting baseline, while the Kalman Filter smoothed RSSI fluctuations to reduce noise. The MLP model was then applied to capture nonlinear patterns in the filtered data and refine position estimates. This integration allowed the system to overcome the weaknesses of traditional methods by balancing stability and predictive accuracy. The hybrid approach achieved an average positioning error of approximately 0.4 meters, which is significantly lower than RADAR. These findings confirm that traditional algorithms like RADAR are limited in handling environmental variability. In contrast, hybrid models incorporating machine learning and filtering techniques adapt better to dynamic indoor settings. The superior accuracy of the hybrid method highlights its suitability for real-time applications where precision is critical. Overall, the comparison underscores the evolution of IPS research from simple fingerprinting approaches to advanced hybrid solutions that achieve robust and highly accurate localization.