Improving localization accuracy based on Lightweight Visual Odometry

New methods based on vision have emerged in the area of mobile vehicle localization. Such methods offer an improved alternative in terms of accuracy to traditional localization methods like wheel odometry. In this paper we propose such a method that does not compromise precision and can run in real time. Depending on environment, feature numbers are sometimes insufficient. To solve this, our algorithm allows using slower feature detectors like SURF for frame keypoints, together with Shi-Tomasi corners for increasing points number. We show how accuracy is further improved by using a Kalman filter to enhance the computation of pose to pose relative motion variation.