Machine Learning Approaches for WiFi Round Trip Time Indoor Positioning Systems

Indoor positioning systems based on WiFi Round-Trip Time (RTT) measurements, as per the IEEE 802.11mc standard, have demonstrated sub-metre level accuracy using trilateration under ideal indoor conditions. However, the efficacy of WiFi RTT positioning in complex, non-line-of-sight (NLOS) environments remains an open research question. Therefore, this thesis addresses the challenge by proposing novel machine learning algorithms and validating their performance through extensive empirical experiments in real-world testbeds.

Recent literature has shown improvements in WiFi fingerprinting systems utilising deep learning methods, achieving sub-metre accuracy. However, it was observed that simpler neural networks can sometimes outperform complex ones in certain environments. Moreover, our comprehensive survey of public WiFi datasets has identified several limitations, all of which pose challenges to accessing or accurately using these datasets over time.

To provide a comprehensive analysis of WiFi RTT for indoor positioning, we investigate its properties in several real-world indoor environments on heterogeneous smartphones. We present three publicly available datasets collected on large-scale real-world scenarios, containing both RTT and received signal strength (RSS) signal measures. Using the proposed datasets, we achieved a baseline accuracy below 0.7 metres.

WiFi RTT has shown promising sub-metre level accuracy under a clear line-of-sight path to the user. However, typical workplace environments often cause wireless signals to reflect, attenuate, and diffract. Identifying the NLOS condition of WiFi Access Points (APs) is thus crucial for indoor positioning systems. To this end, we propose a novel feature selection algorithm for NLOS identification of WiFi APs. Utilising RSS and RTT as inputs, our algorithm employs multi-scale selection and machine learning-based weighting methods to identify the most optimal feature sets, achieving an accuracy of up to 98% in NLOS detection of APs.

Different WiFi technologies and algorithms for indoor positioning have strengths and weaknesses that vary by location. Thus, we propose an algorithm to dynamically switch to the most effective positioning model for an unknown location using a machine learning-based weighted model selection algorithm. We evaluated our algorithm across different complex real-world indoor scenarios, demonstrating an improvement of up to 1.8 metres compared to the standard WiFi fingerprinting technique.