Accurate and fast localization in unstructured environment based on shape context keypoints

The goal of this paper is the localization of a car within an unstructured, outdoor area, based on low level algorithms. This is contrary to the current main focus of many researchers, which mainly choose either highly accurate environment observations, e.g. provided by laser sensors, or information rich vision based localization that requires computational expensive algorithms. The localization is achieved by fusing the information of different, noisy environment sensors, like stereo video or ultrasonic, together into one occupancy grid, identify salient points of interest, augment them with appropriate descriptors, match those descriptors against a built up database and remove the outlier matches with an optimized RANSAC algorithm. The paper proposes a flexible framework with six stages that is used to analyze the influence of the different algorithms per stage. The paper concludes with a recommendation on algorithms that are able to localize the car within an error of 10cm in under 10ms single threaded on a standard central processing unit (CPU), containing all necessary calculation steps starting from the grid map.