IVPR: An Instant Visual Place Recognition Approach Based on Structural Lines in Manhattan World

Abstract

Visual place recognition plays an increasingly important role in robot navigation, autonomous flight of unmanned aerial vehicle (UAV), personal navigation, and so on. However, some texture-less scenarios and computational cost are still challenging. In order to improve the performance of visual place recognition with low computational complexity in the texture-less Manhattan world, this article proposes a fast and reliable method using bag-of-words of structural lines represented by the Extended Line Band Descriptors with Pose estimation capability, namely ELBDP. Furthermore, a robust method that only uses one pair of vanishing points and one structural line to estimate the relative pose between image pairs is proposed. The results show that proposed method outperforms point-based approaches such as oriented features from accelerated segment test and rotated binary robust independent elementary features (ORB)-based place recognition, in terms of efficiency, precision, recall, and accuracy of relative pose estimation.