Several adaptations of maximum likelihood approaches to incremental map learning have been proposed recently. In particular, an incremental network optimizer based on stochastic gradient descent provides a fast and easy-to-implement solution to the problem. In this paper, we illustrate two map builders that process laser scans in order to extract the constraint network for the optimization algorithm. The first algorithm builds a map in the form of a collection of scans corresponding to a subset of the poses of a robot moving in the environment. Even though such a solution has the advantage of simplicity, it requires careful processing of data associations. After a preliminary selection of pose constraints candidates, a relative pose is computed through standard scan matching techniques. The second map builder stores a hybrid metric-topological representation: the map consists of a graph whose nodes contain local occupancy grid maps and whose edges are labeled with the relative pose between pairs of nodes. Each patch map summarizes the information of Consecutive raw scans and such a richer representation better solves loop closure. Association between pairs of local maps is then performed and tested using correlation-based techniques. Our aim is to illustrate the effectiveness of a tree network optimizer integrated with simple methods for data association. Experiments reported in the paper show that a compact system integrating the optimizer and one of two versions of the map builder works reasonably well with commonly used benchmarks.
展开▼
