Neural-network multiple models filter (NMM)-based position estimation system for autonomous vehicles

A highly accurate and reliable vehicle position estimation system is an important component of an autonomous driving system. In generally, a global positioning system (GPS) receiver is employed for the vehicle position estimation of autonomous vehicles. However, a stand-alone GPS does not always provide accurate and reliable information of the vehicle position due to frequent GPS blockages and multipath errors. In order to overcome these problems, a sensor fusion scheme that combines the data from the GPS receiver and several on-board sensors has been studied. In previous researches, a single model filter-based sensor fusion algorithm was used to integrate information from the GPS and on-board sensors. However, an estimate obtained from a single model is difficult to cover the various driving environments, including urban areas, off-road areas, and highways. Thus, a multiple models filter (MMF) has been introduced to address this limitation by adapting multiple models to a wide range of driving conditions. An adaptation of the multiple model is achieved through the use of the model probability. The MMF combines several vehicle models using the model probabilities, which indicate the suitability of the current driving condition. In this paper, we propose a vehicle position estimation algorithm for an autonomous vehicle that is based on a neural network (NN)-based MMF. The model probabilities are determined through the NN. The proposed position estimation system was evaluated through simulations and experiments. The experimental results show that the proposed position estimation algorithm is suitable for application in an autonomous driving system over a wide range of driving conditions.