Improving Low-Cost INS/GPS Positioning Accuracy Using Quantile Regression Forests

GPS and INS are two widely used navigation systems to determine positioning and attitude information for land vehicle navigation. However, standalone GPS and INS have limitations, like signal blockage for GPS and growth of position errors with time for INS. To overcome the shortcomings of standalone GPS and INS systems an integrated INS/GPS navigation system is developed. Thus, INS/GPS integrated system has been widely applied for land vehicle navigation which provides precise navigation and positioning solution by overcoming shortcomings of standalone GPS and INS systems. However, low-cost INS could induce errors that may result in large positional drift. Hence, to improve positioning precision of low-cost INS/GPS integrated navigation system during GPS outages, a novel technique called Quantile Regression Forests (QRF) is proposed. The traditionally employed Kalman Filtering (KF) approach has limitations related to stochastic error models of inertial sensors, immunity to noise, and linearization. To overcome these limitations Artificial Neural Networks (ANN) were introduced. Though, ANNs show better prediction accuracy compared to traditional techniques they suffer from local minimization and overfitting problems. To overcome the inadequacies of existing methods, statistical learning techniques like Random Forest Regression (RFR) based on Decision Trees (DT) were implemented. The accuracy of RFR is limited as its predictions are based on averaging the conditional mean of all the observations. Thus, to improve the prediction accuracy of RFR, the paper introduces QRF whose predictions are based on conditional quantiles. The performance of QRF is verified using real field test data and is also compared to both conventional ANN and recently published RFR.