BeaconZone Blog

New research presents RELoc, a WiFi fingerprinting indoor localisation framework designed to work reliably in multi-floor buildings where conventional 2D approaches often struggle because they cannot properly resolve vertical (between-floor) ambiguity. The method combines Recursive Feature Elimination with Cross-Validation (RFECV) to select the most informative WiFi access point signals, and an Extremely Randomised Trees regressor to predict positions as either 2D coordinates or full 3D coordinates including floor information. The trees model is tuned using Bayesian hyperparameter optimisation via Optuna’s Tree-structured Parzen Estimator, with the aim of improving accuracy while keeping computation manageable for practical deployments.

The authors evaluate RELoc on two public datasets, SODIndoorLoc and UTSIndoorLoc, and report that the 2D version achieves mean absolute errors of 1.84 m (SODIndoorLoc) and 4.39 m (UTSIndoorLoc). When floor level is incorporated for 3D prediction, performance improves markedly compared with the corresponding 2D setup, reducing error by about a third on SODIndoorLoc and by just over a quarter on UTSIndoorLoc, which the paper attributes to 3D modelling’s ability to separate locations that look similar in WiFi signal strength in the horizontal plane but lie on different floors. Across both datasets, the reported results show RELoc outperforming a range of baseline machine learning, ensemble, and deep learning methods, while also training faster than heavier neural approaches in their experiments.

The paper concludes that combining cross-validated feature selection with an efficient tree ensemble and automated tuning produces a strong balance of accuracy and computational efficiency for multi-floor indoor positioning.

A new paper addresses a core limitation of beacon-based indoor localisation systems, unstable and noisy RSSI measurements that degrade distance estimation and, by extension, location accuracy. The work focuses on Bluetooth Low Energy beacon deployments as a primary use case, alongside Wi-Fi and Zigbee, within Indoor Spatial Temporal Systems that rely on trilateration from RSSI values .

In Bluetooth beacon systems, RSSI values fluctuate significantly indoors due to multipath effects, attenuation from walls and furniture, human movement and device interference. The paper shows that conventional approaches, where RSSI filtering is performed in the cloud using Kalman or adaptive Kalman filters, introduce unacceptable latency for real-time beacon applications such as indoor navigation, proximity detection and safety monitoring. Moving filtering closer to the beacons at the edge reduces transmission delay but introduces a new problem. Adaptive filters such as robust self-adaptive Kalman filters are computationally expensive and unsuitable for resource-constrained edge devices commonly used with Bluetooth beacons .

To address this, the authors propose a lightweight edge-Kalman Filter designed specifically for noisy RSSI streams like those produced by Bluetooth beacons. Instead of continuously updating noise parameters, the filter only updates when a statistically significant change is detected between consecutive RSSI windows using density ratio estimation. This reduces unnecessary computation while still responding to real environmental changes, making it well suited to beacon receivers such as Raspberry Pis or mobile devices operating at the edge .

Experimental results using multiple Bluetooth beacon datasets show that the proposed approach substantially reduces distance estimation error compared with raw RSSI, standard Kalman filtering and robust self-adaptive Kalman filtering. In beacon scenarios with real-world noise, the edge-based approach achieves lower mean squared error while requiring fewer computations, which directly improves responsiveness and quality of service for Bluetooth beacon applications. The paper also demonstrates that cleaner, filtered beacon RSSI significantly improves downstream machine learning models for indoor location prediction, increasing classification accuracy to near-perfect levels in tested scenarios .