A Probabilistic Graphic Model based GPS/SINS Integration Algorithm for Low-cost Sensors on Smartphone

Abstract

With the popularity of smartphones and the development of navigation sensor technology, smartphones are expected to serve as unified platform for ubiquitous location services. In land vehicle navigation application, a GPS receiver equipped smartphone can be used for several location-based services. However, short-term GPS outage may occur in urban environment when crossing tunnels or viaducts. An ideal solution could be a combination of GPS and low cost inertial measurement sensors using GPS/SINS integrated navigation algorithm. Traditional GPS/SINS integrated navigation algorithms are based on Kalman type filter with Strap-down Inertial Navigation error propagation model and inertial sensor error model as the system model. However the high uncertainties of nonlinear drift in low cost Micro-Electro-Mechanical System (MEMS) Inertial Measurement sensors make the errors are difficult to characterize. It is still a challenge to develop an optimal real-time integration algorithm that can maintain required system performance during GPS short-term outage (maximum 30s) for low cost sensors on smartphone. In this paper, we proposed a reliable hybrid positioning method by combining the advantages of belief propagation (BP) and structure learning in the probabilistic graphical model, which addresses GPS outages and uncertain nonlinear drift of MEMS INS simultaneously. The state estimation problem of integrated navigation was formulated in the framework of factor graph. A factor graph is a probabilistic graphical model which combines the graph theory and probability theory to give a multivariate statistical modeling. An unscented transform based belief propagation algorithm was proposed to address multi-rate measurement and partial nonlinear system model. This factor graph based state estimation was augmented by another adaptive probabilistic graphical model which was used to model position error induced by MEMS IMU sensors nonlinear drift. This architecture works in learning mode when GPS measurement is available and works in prediction mode when GPS measurement is in short-term outage. A road-test experiment in a land vehicle was taken in order to verify the method proposed in this paper.