Indoor Localization: BLE, Machine Learning, and Kalman Filtering for Resource-Constrained Devices

Abstract

The growing complexity of indoor environments and limitations of the Global Positioning System (GPS) require accessible yet efficient indoor localization solutions in areas such as navigation and asset management across various sectors such as airports, and warehouses. Existing solutions often come with a trade-off between accuracy, efficiency, and setup burden, hindering their deployment on resource-constrained edge devices.

This project addresses the challenge by introducing a novel indoor localization system that leverages Bluetooth Low Energy (BLE) beacons and lightweight machine learning models such as the Random Forest and KNN to develop a system capable of accurate and computationally efficient location prediction. By combining Received Signal Strength Indicator (RSSI) fingerprinting approaches with preprocessing techniques, resulting data can be cleaner and mitigate major issues caused by sensor noise and multipath fading effects. This preprocessing step allows for faster model training and facilitates deployment on low-power edge devices like mobile phones, robots, and microcontrollers without the need for more complex solutions.

This project offers a comprehensive evaluation of varying combinations of preprocessing algorithms and machine learning models. The results demonstrate that the Kalman filter can significantly reduce multipath and sensor noise, enabling faster model training and higher accuracy. The Kalman filter allowed the Random Forest model to achieve an accuracy of 95% given just 52 training samples, resulting in prediction times under 5 microseconds. This eliminates the need for complex, resource-intensive solutions, empowering the system to achieve high performance with efficient machine learning.