Smartphone Sensor Data Fusing For Indoor Localization

The smartphone has multiple sensors that senses information about the indoor location.Fully integrating multiple sensors data has an important role in improving indoor localization accuracy,and has become a hot issue in recent years.The phone's magnetometer can sense the intensity of the geomagnetic field.The geomagnetic field is ubiquitous and stable.Local distortion occurs due to the indoor magnetic substances.These distortions allow the geomagnetic intensity to distinguish between different indoor locations.WiFi AP(Access Point)is widely deployed indoors.WiFi's Received Signal Strength(RSS)is attenuated with the transmission distance and can also be used for indoor localization.However,how to make full use of the characteristics of both to designing low-cost and high-precision indoor localization algorithms is still a challenges.Through experimental research,it is found that the geomagnetic field and the WiFi signal have complementary characteristics in terms of location information granularity and location discrimination.Based on this,a two-stage localization framework for smartphones is proposed: coarse-grained localization stage and fine-grained localization stage.Two localization algorithms based on geomagnetic and WiFi fusion are implemented under this framework: MagWi and DeepLoc.MagWi is a feature fusion indoor localization algorithm based on error distribution weights.It uses two-dimensional geomagnetic features that are independent of the phone's attitude.In order to make full use of the complementarity of geomagnetism and WiFi,MagWi has established a prediction error model that can describe the localization capabilities of features.MagWi uses the WiFi RSS feature to determine a sub-area in coarse-grained localization.In the fine-grained localization phase,the two-dimensional geomagnetic feature and the WiFi feature are dynamically weighted and integrated based on the prediction error model to obtain the user location.DeepLoc is a deep learning based localization algorithm.It divides the location area into multiple sub-areas in the offline phase,and trains a DNN(Deep Neural Network)classification model for sub-area classification,and then trains an independent DNN regression model for each sub-area for location estimation.Geomagnetic and WiFi data are collected in real time during the online phase,and the trained DNN model is used to locate the user.DeepLoc implements two stages of deep fusion optimization for geomagnetism and WiFi.Both MagWi and DeepLoc use only the sensors that come with commercial phones,eliminating the need for additional hardware modules and the need to deploy dedicated facilities in advance,so deployment costs are low.This paper implements MagWi and DeepLoc on Android smartphones and servers,and performs experimental verification in actual scenarios.The experimental scene consists of two teaching building corridors,a conference room and a library hall with a total area of over 700 square meters.Experiments show that MagWi and DeepLoc have their own advantages and disadvantages in different scenarios,but they can achieve higher localization accuracy than traditional methods,and can meet the low-cost and high-precision localization requirements.