Visual Place Recognition for Opposite Viewpoints and Environment Changes

Human beings possess a remarkable capability for visual scene understanding due to the inherent hardwiring of our brain’s visual system. This allows us to effortlessly recognize a place even when approaching it from the opposite direction, without the need to physically turn around. However, replicating this ability in autonomous vehicles or robots, a field known as visual place recognition (VPR) presents significant challenges. In the case of front–front VPR, the primary challenges revolve around changes in appearance and lighting conditions. However, for front–rear VPR, an additional complex issue arises, as a substantial portion of the visual field does not overlap between two opposing scenes. In this research article, we propose an innovative pipeline for addressing the front–rear VPR problem, using the vector of locally aggregated descriptors (VLAD). In the coarse matching stage, we use a combination of two complementary descriptors. One descriptor captures the scene’s semantics, while the other calculates the utility of clusters based on appearance. We then use clusters with high utility to construct a novel local image descriptor known as L-NetVLAD. In addition, we use a sequence matching method for contrastive learning on these descriptors to compute distances and optimize their performance. For the fine matching stage, we implement a double-filtering approach for keypoints, using available semantic information to further refine and eliminate incorrect keypoint correspondences. Our experiments on various benchmark datasets demonstrate the robustness and high accuracy of our method in tackling all the aforementioned challenges.