Three-dimensional localization and mapping of static environments by means of mobile perception

Abstract

Model-based task planning is one of the main capabilities of autonomous mobile robots. Especially for model-based localization and path planning, a large-scale description of the operation environment is required. Cognitive communication between man and his machine could be based on a common, three-dimensional understanding of the environment. In the case of a personal service robot, the operation environment may comprise both indoor and outdoor spaces. In this thesis, a method for the generation of a three-dimensional geometric model for large scale, structured and natural environments is presented.

The environment mapping method, which uses range images as measurement data, consists of three main phases: first, geometric features are extracted from each of the range images. Secondly, the relative coordinate transformations (i.e. registrations) between the sensor viewpoint locations, where the range data was measured, are computed. And, finally, an integrated map is formed by transforming the sub-map data into a common frame of reference.

Two types of geometric features are extracted from the range images: cylinder segments (or more generally truncated cone segments) and straight-line segments. With cylinder segments tree trunks and other elongated cylindrical objects can be modeled, whereas the straight line segments correspond to the upper corners of vertical walls. The features are utilized as natural landmarks for registration computation.

The presented method is tested by mapping three test sites representing structured, semi-structured and natural environments. The structured environment corresponds to a part of the premises of an office building, the semi-structured environment corresponds to the surroundings of a parking lot and the natural environment is a small forest area. The dimensions of the test sites are about 50 meters, 120 meters and 40 meters square, respectively. A simple incremental approach is used to build an integrated model for the parking lot and office corridor environments. For the principal mapping experiment, concerning the small forest area, a statistically more sound, optimal approach is applied. With respect to the feature extraction methods and the computation of the relative coordinate transformations between the viewpoints, robustness to outlier data and failure modes of the methods are discussed in more detail.