Ensemble of Bayesian Filters for Loop Closure Detection

Loop closure detection for visual only simultaneous localization and mapping needs effective feature descriptors to obtain good performance results. Currently, the most widely used feature description is the global or local descriptor such as color histogram and Speeded Up Robust Features. The global features can be computed either by considering all points within a region, or only for those points on the boundary of a region. In contrast, the local features are obtained by considering the boundary of an object that represents a distinguishable small part of a region. One possible problem of these approaches is that the number of features become very large when a dense grid is used where the histograms are computed and combined for many different regions or points. The most popular solution for the problem is to use a clustering algorithm to create a visual codebook to create a histogram of visual keywords present in a visual image. In this paper, we designed and implemented an ensemble learning method namely mean rule to combining three different local features: Scale-invariant feature transform (SIFT), Speeded Up Robust Features (SURF) and Oriented FAST and Rotated BRIEF (ORB). The aim of using ensemble learning is to enhance learning speed and final performance of different local visual keywords descriptors for loop closure detection. Furthermore, the Real-Time Appearance-Based Mapping (RTAB-Map) using a Bayes filter is used to evaluate loop closure hypotheses. Experimental results on a public dataset contains 2464 images show that the ensemble algorithm outperform the single bag-offeatures approach.