A Graph Approach to Dynamic Fusion of Sensors

摘要

PGM-based or PGM-inspired simultaneous localization and mapping (SLAM) has been successfully adopted for both online and offline pose estimation and map construction using multiple sensor modalities. One particular problem of this graph framework is the computational complexity: the longer the experiment, the larger the problem to solve; the more the sensor modalities, the larger the problem to solve. Several improvements have been made to constrain the time complexity. In this work, an information-theoretic method is applied to prune/simplify the graph while maintaining the a posteriori information represented by the underlying graph. In particular, Chow-Liu tree is used to connect the nodes of the Markov blanket of the node to be eliminated and the factors of the graph is reconstructed by computing mutual information between each and every pair of the nodes within the blanket and forming the network topology in a maximum spanning tree (MST). Synthetic results show that the pruned and reconstructed graph achieves similar estimation accuracy compared with the original graph, which is clearly signified by the Kullback-Leibler divergence metric.