Dynamic Modeling for Land Mobile Navigation Using Low-Cost Inertial Sensors and Least Squares Support Vector Machine Learning

K. Frangos, A. Kealy, V. Gikas, A. Hasnur

Abstract:	Traditional algorithms used to determine a vehicle’s navigation state (e.g. Kalman filter) has as one of its prerequisites, a model that describes how the vehicle is expected to move over time. The accuracy of this dynamic model is important, as it allows for optimization of the navigation solution, particularly when dealing with low cost sensors which typically exhibit significant errors and biases. Unfortunately, for land vehicles, apriori knowledge of the true dynamic model is very difficult to achieve by virtue of the random dynamic variations that exist and that there is no general navigation case. This situation is even further complicated in many navigation applications where non-linearity and demanding environments characterize the motion and challenge the assumptions of most filtering methods (e.g. linear dynamics behaviour and Gaussian white noise). To overcome these problems, a new approach has been developed to determine the correct dynamic model for the navigation platform; this approach is based on Support Vector Machines (SVM). SVM is a relatively new learning technique in the field of machine learning and is based on structural risk minimization (SRM) which makes the technique statistically robust. SVM addresses and overcomes the majority of problems faced by typical neural networks such as local minima, over-fitting or over-training, etc. In this research, Least Squares Support Vector Machine (LS-SVM); a sub case of the SVM theory is utilised as a means of maintaining the ratio between overall performance and computational overheads. The approach taken here is to identify the true system dynamics by learning from a time series analysis of a set of training data. A mathematical model which describes the true system dynamics regression is then created from this analysis. This model can then be applied to predicting the behaviour of the navigation platform. In terms of filtering techniques, this is the first step for the formulation of an adaptive dynamic model for state prediction, replacing the standard Kalman filter´s transition phase. In this paper, the LS-SVM algorithm used in this research for dynamic modeling is detailed. In addition, practical results that describe the performance of this algorithm will be presented. These results have been generated using a navigation test bed established in Melbourne, Australia. The test data was captured on a land vehicle fitted with one tactical grade IMU, six low cost MEMS IMU sensors and an array of high performance dual frequency GPS receivers. The realworld tests, data captured, analysis, processing techniques and dynamic modeling results will be described and used to demonstrate the performance of the LS-SVM algorithm.
Published in:	Proceedings of the 23rd International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS 2010) September 21 - 24, 2010 Oregon Convention Center, Portland, Oregon Portland, OR
Pages:	1687 - 1696
Cite this article:	Frangos, K., Kealy, A., Gikas, V., Hasnur, A., "Dynamic Modeling for Land Mobile Navigation Using Low-Cost Inertial Sensors and Least Squares Support Vector Machine Learning," Proceedings of the 23rd International Technical Meeting of the Satellite Division of The Institute of Navigation (ION GNSS 2010), Portland, OR, September 2010, pp. 1687-1696.
Full Paper:	ION Members/Non-Members: 1 Download Credit Sign In