This thesis introduces three important components of autonomous navigation: visual odometry and image fusion, Kalman filtering and its application, simultaneous localization and mapping (SLAM). And presents Fast - SLAM 3.0: an improved approach to SLAM compared to Fast SLAM 2.0 and Extended Kalman Filter (EKF) SLAM. The Fast SLAM 3.0 models the particles as the robot pose mean of a Gaussian distribution, which keeps the error covariance matrix (P) of pose estimation propagating as normal EKF SLAM. Hence uncertainty could be remembered over the whole trajectories, avoiding Fast SLAM 2.0's tendency to become over-confident and keeping the best feature of Fast SLAM that locally avoids linearization of the robot model and provides a high level of robustness to the clutter and ambiguous data association. Extensive experiments in the randomly generated simulated environment show that the Fast SLAM 3.0 significantly outperforms either Fast SLAM 2.0 or EKF SLAM.