Fusion Of Smartphone Sensors For Indoor Pedestrian Trajectory Recovery:A Data-Driven Approach

With the popularity of smartphones equipped with rich sensors,smartphones and a series of mobile terminals are evolving into new sensing and computing platforms.By virtue of the convenience of obtaining location information from smartphones,location-based service has attracted more and more attention from industry.The key is to obtain accurate and reliable indoor pedestrian trajectories.The state-of-the-art technologies have been able to use sensory data from smartphones to determine the holder's footsteps and walking direction,and can obtain a rough estimate of the real position based on wireless signal strength scanned by smartphones,but these methods will inevitably suffer from noise and accumulated estimation errors,as well as the effects of signal fluctuations,which affect the estimation accuracy.In addition,the current high-precision indoor positioning and navigation techniques mostly rely on specific indoor arrangements and the deployment of special devices,such as the radio frequency identification,infrared rays,visible light,computer vision,and ultrasound,etc.Although these methods can provide accurate position estimates,they are difficult to be popularized into daily life and are very expensive.In order to achieve high trajectory recovery accuracy in a general indoor environment,a promising way is to combine the existing techniques,to enhance the modeling capacity,and to learn the complex dynamics from accumulated historical data.In this thesis,a data-driven indoor trajectory recovery system is proposed.The non-parametric Gaussian process is used to model the mapping between the position of a pedestrian on two consecutive steps,as well as the mapping between the ground-truth position and related measurements.The Gaussian process state-space model(GPSSM)can more accurately capture the dynamics of a walking pedestrian.Compared with the parametric and empirical models commonly used in indoor scenarios,data-driven and non-parametric models can better represent the complex dynamics and the measurement data collected in wireless environments.The GPSSM provides a natural fusion mechanism for integrating the cheap received-signal-strength based WiFi localization technique and the pedestrian dead reckoning approach,which leverages smartphone built-in inertial measurement units to perform step detection and walking direction estimation.In addition,compared with the “black box”-type deep neural networks,the Bayesian method represented by Gaussian process has better model interpretability and flexibility.Facing the special and complex structure of the GPSSM,this thesis combines the latest research outputs of Bayesian inference and deep learning on designing an accurate hyperparameter training framework with low computational complexity,which can be easily extended to large data sets.This thesis also implements the GPSSM for indoor trajectory recovery and proposes a customized optimization scheme with less computational complexity and more stability.A software system for smartphone sensory data collection,transmission,storage,and management is designed and developed,which is used to collect trajectory related data in a real and complex indoor environment.The accuracy of the recovered trajectory brought by the GPSSM as well as the model's excellent adaptability are demonstrated in this thesis.