Using Machine Learning to Improve Accuracy and Robustness of Indoor Positioning under Practical Usage Scenarios

Indoor Positioning Systems (IPSs) can increase productivity in both office and industrial settings. They continue to become more accurate and robust as the advent of machine learning enables them to overcome the limitations of traditional positioning techniques. Despite this, the mainstream incorporation of IPS is currently hindered by significant infrastructure cost, especially for areas that cannot attain sufficient wireless coverage due to budget or environmental constraints. This paper therefore explores the use of machine learning for infrastructure-limited smartphone-based localization while adhering to practical constraints. The performance of the trained models was compared to that of conventional multilateration while also considering the effect of phone placement on positioning accuracy. Experimental results showed that the model trained under harsher conditions proved to be the most robust for both handheld and pocket mobile tests.