Low-cost loosely-coupled GPS/odometer fusion: a pattern recognition aided approach

By combining statistical pattern recognition techniques and a classic multiple-model vehicle tracking framework, the presented algorithm improves the performance of fusing GPS outputs and vehicle motion data. The multiple-model tracking framework is used to represent vehicle motion as a combination of several maneuvers, whilst, pattern recognition techniques are proposed to identify the model that corresponds to the true vehicle maneuver. This recognition approach does not use any filter states, therefore it is robust to errors in Kalman filter modeling. To train the classifier, several data collection trials were performed to acquire GPS outputs and vehicle responses in urban traffic conditions. This approach is designed for low cost GPS receivers and data available in modern passenger cars. Extensive experiments were conducted in Paris region in real traffic conditions. The performance is benchmarked with respect to a ground-truth trajectory generated by a high-end GPS-Inertial-Odometer based localization system. Promising results were observed and a discussion of them is included.