Ensemble of Vector and Binary Descriptor for Loop Closure Detection

Loop closure detection plays an important role in vSLAM for building and updating maps of the surrounding environment. An efficient vSLAM system needs an informative descriptor for landmark description and stable model for making decisions. Most of the solutions dependent on using a single descriptor for landmark description, whereas other solutions proposed to use a combination of descriptors. However, these solutions still have the limitation in correctly detecting a previously visited landmark. In this paper, an ensemble of loop closure detection is proposed using Bayesian filter models for making decisions. In this approach, a set of different keypoint descriptors is used as input to bag-of-word descriptors. After that, these descriptors, i.e., SIFT, SURF, and ORB, are used to construct Bayesian filter models and ensemble learning algorithm for loop closure detection. The proposed approach is validated on a public dataset, namely City-Center dataset (CiC). The results shown that the proposed ensemble algorithm outperforms single model and existing loop closure detection system approaches. It gives 87.96 % for ensemble learning and 86.36 % for the best single model and 37, 80, 81 % for FAB-MAP, PIRF-Nav2.0, and RTAB-MAP, respectively.