Delayed resampling in a Rao-Blackwellized particle filtering SLAM for consistent loop closures

The Rao-Blackwellized particle filtering SLAM (RBPF-SLAM) still suffers from a large-loop or nested-loop closure due to a lack of hypotheses. This makes the filter optimistic and results in the RBPF to be a short term solution. This paper provides an important insight to the current resampling scheme and proposes an approach to delaying the resampling decision until the loop closing. A discussion about the issue of preserving the importance weight is also presented with developing the weight-smoothing method. This approach is a necessary step for the RBPF-SLAM, but to the best of our knowledge, no literature deals with this concept. Since the delayed resampling maintains the possible hypotheses, the filter consistency is significantly improved. Moreover since the resampling takes the loop closure into account, the required number of particles can be proportional to the loop size of the environment. The computational burden, caused by the resampling process, is also drastically reduced due to the delaying decision. Experimental results in large-scale environment with large and nested loops illustrate the effectiveness of our approach.