SPOT: Point Cloud Based Stereo Visual Place Recognition for Similar and Opposing Viewpoints

TL;DR

SPOT addresses opposing-viewpoint place recognition with limited-FOV stereo cameras by leveraging structure from stereo visual odometry. It builds Cart Context descriptors from equi-spaced keyframe point clouds and applies a novel double distance matrix sequence matching, including a query-side double-flip to handle opposing viewpoints without expanding the reference database. On the NSAVP dataset, SPOT achieves up to 91.7% recall at 100% precision for opposing viewpoints and requires less storage while running efficiently on CPU, outperforming state-of-the-art baselines. The approach demonstrates the viability of VO-derived structure for robust VPR and is well-suited for integration into SLAM systems and cross-modality place recognition scenarios.

Abstract

Recognizing places from an opposing viewpoint during a return trip is a common experience for human drivers. However, the analogous robotics capability, visual place recognition (VPR) with limited field of view cameras under 180 degree rotations, has proven to be challenging to achieve. To address this problem, this paper presents Same Place Opposing Trajectory (SPOT), a technique for opposing viewpoint VPR that relies exclusively on structure estimated through stereo visual odometry (VO). The method extends recent advances in lidar descriptors and utilizes a novel double (similar and opposing) distance matrix sequence matching method. We evaluate SPOT on a publicly available dataset with 6.7-7.6 km routes driven in similar and opposing directions under various lighting conditions. The proposed algorithm demonstrates remarkable improvement over the state-of-the-art, achieving up to 91.7% recall at 100% precision in opposing viewpoint cases, while requiring less storage than all baselines tested and running faster than all but one. Moreover, the proposed method assumes no a priori knowledge of whether the viewpoint is similar or opposing, and also demonstrates competitive performance in similar viewpoint cases.

Figures (6)

• Figure 1: Overview of SPOT. External inputs are green, processing blocks are blue, outputs are white. The stereo VO algorithm, SO-DSO mo_extending_2019 , processes stereo images and outputs estimated poses and scaled depth images. This output is accumulated to form a point cloud at a selected keyframe pose. Next, a Cart Context kim_scan_2022 query descriptor is formed. Two distances between the query and all references are computed and a novel double distance matrix sequence matching scheme produces the final reference match.
• Figure 2: An expanded depiction of the final three place recognition stages. External inputs are green, processing blocks are blue, outputs are white, and incrementally updated objects are orange. A Cart Context kim_scan_2022 descriptor is formed from the most recent keyframe point cloud and flipped about both axes to produce an additional descriptor for opposing viewpoint VPR. Descriptor distances are computed between each query and every reference to produce two separate distance matrices for similar and opposing viewpoints. Sequence matching, as described in milford_seqslam_2012, is performed separately in each distance matrix and the final output is selected from the results.
• Figure 3: A depiction, for a single longitudinal ($k$) and lateral ($l$) shift, of the overlapping regions of the query $\mathbf{Q}$ and reference $\mathbf{R}$ between which the cosine distance is computed.
• Figure 4: Opposing viewpoint images from route R0nsavp at the same place. Left to right: noon reference, noon query, sunset query and night query.
• Figure 5: Evaluated routes from the NSAVP dataset.