Artificial neural network-based indoor localization system using smartphone magnetometer

International audienceIn recent years, localization systems have becomean interesting research topic since they present a key factorfor location-based services. Although the global positioningsystem (GPS) is widely used for outdoor positioning, it doesnot provide the same accuracy in indoor environments. As aresult, many alternative indoor positioning technologies havebeen investigated to tackle this problem during the last fewyears. However, most existing approaches (e.g., Camera, WiFi,and infrared-based methods) for indoor localization mainly relyon infrastructure, which is expensive and not scalable. Today,the expansion of smartphones possessing a variety of embeddedsensors helped develop a precise indoor localization that canmeet the requirements of location-based services.This paper proposes an indoor localization system based onmagnetic field sensed via a smartphone magnetometer. Anomaliescaused by the presence of ferromagnetic materials are usedas signatures and fingerprinting to identify different locations.Accordingly, an Android application was developed to build ageomagnetic fingerprinting database in the corridor of PolytechNantes, France. 7600 signatures were stored in the database,cleaned, and standardized. 70% of data is used to train andvalidate different multi-output regression models. Extensivesimulations are conducted to find the suitable model and totune model’s hyper-parameters. Once the model is configuredand trained, remaining unseen data is used to evaluate theaccuracy of the proposed system. Obtained results demonstratethat ANN is the most accurate model with Mean Absolute Error(MAE) equals 0.13m for the studied environment. Only 8% oftesting samples have errors higher than the MAE. Moreover, theproposed indoor localization system can locate the user withoutprior knowledge of his initial position