Abstract
Indoor positioning systems (IPS), one of the emerging technologies, play an important role in the Internet of Things (IoT) that addresses the challenges of GPS in indoor environments. This technology creates a great impact on real world application. Bluetooth Low Energy (BLE) has emerged as one of the most cost-effective, energy-efficient mechanisms for facilitating indoor localization. A problem with many available BLE fingerprinting techniques is that they do not consider differences in signal characteristics due to the orientation of the device, and limited statistical evidence exists to demonstrate variations in performance. The present study developed a machine learning-based indoor localization framework using BLE 5.1 Received Signal Strength (RSS) and Angle-of-Arrival (AoA) data to evaluate the effect of a device’s orientation on decimeter-level localization accuracy. Real-world static datasets were assembled under four orientation conditions: North, South, East, and West. In addition, temporal windowing and median-based fingerprint generation techniques were used to alleviate the impact of signal instability on accuracy. Random Forest (RF), k-Nearest Neighbour (KNN), and Multilayer Perceptron (MLP) were used to generate estimates of 2D positions based on the generated fingerprint. Performance of models was evaluated using distance-based error metrics, percentiles, reliability analysis, and the Wilcoxon signed-rank test for statistical assessment. The results obtained indicate that device orientation is a significant factor affecting localization accuracy, and Random Forest (RF) produces more robust localization accuracy than KNN and MLP across each of the four cardinal orientations. The results also illustrated the need to use orientation-aware modeling techniques to create a reliable BLE-based IPS.
Keywords: Fingerprint Approach, Indoor Positioning System, Median Filtering, Orientation Wise Analysis, Static Scenario.