Bayesian Machine Learning in INS/WiFi Integrated Navigation Systems for Indoor and GNSS-denied Environments

摘要

GPS can provide sub meter accuracy in open-sky areas where the GPS signal is strong and at least 4 satellites are visible by the receiver with good geometry. However, for indoor and dense urban environments, the accuracy deteriorates significantly due to weak signals and dense multipath. The situation becomes worse in indoor environments where the GPS signals are unreliable or totally blocked. On the other side, Inertial Navigation Systems (INS) are self-contained systems that can be used indoor/outdoor evenly. However, INS provides only good short-term accuracy. The mathematical integration causes even small sensor errors to accumulate resulting in large position drifts that grow over time. In addition, if low cost MEMS-based inertial sensors are to be used, then sensors random errors are hard to model because of the high non-linearity in its behavior. Thus, an accurate and reliable indoor positioning is still one of the greatest challenges in the field of navigation. Therefore, in this research we propose an alternative indoor 3D integrated navigation system for wheeled vehicles using the existing IEEE 802.11 WLAN (WiFi) which is widely available in indoor environments, and Low Cost MEMS-based reduced inertial sensors system (RISS). The system development and integration is based on Bayesian Inference techniques; Gaussian Process Regression (GPR) for wireless propagation modeling and optimized adaptive version of Mixture Particle Filter (PF) algorithm for state estimation. Real experiments on a mobile robot showed that the integration between WiFi and RISS improved the overall accuracy of the system providing sub-meter accuracy for 50% of the time, 2.5 meters accuracy for 65% of time and maximum position error of 5.8 meters.