Localizing while mapping: a segment approach

Andrew J. Martignoni,W. Smart

2002-07-28

Abstract:1999; Rencken 1993).Map building with occupancy grids (probabilistic finite el-ement maps of the environment) is also fairly well under-stood. However, trying to accomplish both localization andmapping at once has proven to be a difficult task. Updatinga map with new sensor information before determining thecorrect adjustment for the starting point and heading can bedisastrous, since it changes parts of the map which were notmeant to be changed, thus the next iteration is using an in-correct map to determine the adjustment to the pose, whichresults in greater error, and the cycle continues.This research hopes to solve this problem by perform-ing localization on higher-level objects computed from theraw sensor readings. Initially these higher-level objects areline segments and common structures, such as office doors,butany object which can be detected from a single sensorreading can be used to help the localization process. Ob-jects which require more complex analysis can be found andplaced in the map, but will not be used for localization. Thehigher-level features provide for better localization in an un-known environment, through simpler correspondence withexisting map objects. In addition, once the data from therobot’s sensors have been turned into a map with human-identifiable features, the map can easily be augmented withadditional information.Much of the previous work in this area uses occupancygrids (Moravec & Elfes 1985; Elfes 1987), or pixel basedmaps. Pixels in occupancy grids represent a certain size areaof the environment, and represents the probability that thearea contains an object. Other work (Park & Kender 1995;Kuipers & Byun 1991) uses a topological system wherelandmarks are picked out of the environment and added asnodes on a graph which is then used for navigation.Pixel-based maps tend to limit the size of the area thatcan be covered, and the actual data in the image tends tobe sparse because the map representations are a fixed sizeand shape and do not necessarily match the size and shapeof the area to be mapped. An object map (of line segmentsor other detected features) has no such bounds since only theend points of a segment (or similar real-valued data for other