Extended Kalman Filter in Autonomous Mobile Robot Localization and Mapping

Abstract

The implementation of neural extended Kalman filter is achieved in terms of the monocular SLAM problem: multi-layer perceptron neural network is coupled with EKF to improve the state transition model. The main advantage of NEKF is the ability of the neural network to learn a model of the system on-line. This article showed that the introduction of neural network has resulted in higher accuracy of NEKF than a “standard” extended Kalman filter implementation for monocular SLAM. Multiple repetitions of experiment are performed, and experimental results indicate that NEKF outperforms EKF and odometry in terms of accuracy. Future work could be extended through implementation of other Gaussian filters (Unscented Kalman Filter or Extended Information Filter) and different types of feedforward neural networks (Radial Basis Function or maybe even Hyper Basis Function). To achieve real time performance of 30 [Hz] additional hardware is necessary. Finally, through experiments with the mobile robot and simple USB web camera, we have showed that this approach can be applied but we are looking forward to implement it in the real manufacturing environment and assess its performance.