Efficient Data Association Approach to Simultaneous Localization and Map Building

A feature based approach to simultaneous localization and map building (SLAM) is to use the information obtained by sensors mounted on a vehicle to build and update a map of the environment and compute the vehicle location in that map. One of the critical problems in obtaining a robust SLAM solution is data association, i.e. relating sensor measurements to features in the map that has been built thus far (Guivant, 2000). SLAM relies on correct correspondence between data obtained from the robot sensors and the data currently stored in the map. There have been numerous approaches to data association. In stochastic mapping, the simplest method is the NN algorithm which is a classical technique in tracking problems (Bar-shalom & Fortmann, 1988). The great advantage of NN is its o(mn) computational complexity in addition to its conceptual simplicity (Here m is the number of sensor measurements and n is the number of existing features in the map). It performs satisfactorily when clutter density is low and sensor accuracy is high. However, during the process of SLAM, especially in complex outdoor environments, clutter level is usually high and the innovations in matching different observations obtained from the same vehicle position are correlated. In this situation, the NN algorithm may accept a wrong matching, which leads to divergence in state estimation. In order to improve the robustness of data association, Neira and Tardos (Neira & Tardos, 2001) presented an approach using a joint compatibility test based on the branch and bound search with a high computational cost. Juan Nieto et al. (Nieto et al., 2003) give a fast SLAM algorithm for data association by applying the multiple hypotheses tracking method in a variety of outdoor environments. The experimental complexity estimates show that if the number of features in one scan is large, these algorithms will not be fast enough for real time implementation. In other approaches, Bailey et al. consider relative distances and angles between points and lines in two laser scans and use graph theory to find the largest number of compatible pairings between the measurements and existing features (Bailey et al., 2000). The work of Lim and Leonard (Leonard & Lim, 2000) applies a hypotheses test to implement data association of the relocation in SLAM using geometric constraints. Castellanos and Tardos (Castellanos et al., 1999) use binary constraints to localize the robot with an a priori map using an interpretation tree. In these methods, geometric constraints among features are used to obtain hypotheses with pairwise compatible parings. However, pairwise compatibility doesn't guarantee joint compatibility, and additional validations are required.