OpenGV 2.0: Motion prior-assisted calibration and SLAM with vehicle-mounted surround-view systems

TL;DR

This work tackles localization and mapping with vehicle-mounted surround-view cameras that have limited field-of-view overlap by introducing motion-prior-based modules for online extrinsic calibration, stable motion initialization, and non-holonomic trajectory optimization. It integrates these modules into a full surround-view SLAM system that uses continuous-time splines (B-splines) to enforce Ackermann-like vehicle motion without relying on circular-arc simplifications, and demonstrates robust performance on urban datasets, with optional weak GPS aiding long-term drift control. Key contributions include online camera-to-vehicle orientation calibration using forward/upward directional constraints and vertical-structure cues, a robust multi-camera motion initialization solver, and a spline-based back-end with several variants (CBA, CBARt, CBASpRv, SSBARv, FSBA) that enforce non-holonomic constraints. The approach achieves accurate, drift-resistant trajectory estimates on large-scale public datasets and is slated for open-source release as an extension of the OpenGV library, offering practical impact for low-overhead surround-view perception in urban autonomous driving contexts.

Abstract

The present paper proposes optimization-based solutions to visual SLAM with a vehicle-mounted surround-view camera system. Owing to their original use-case, such systems often only contain a single camera facing into either direction and very limited overlap between fields of view. Our novelty consist of three optimization modules targeting at practical online calibration of exterior orientations from simple two-view geometry, reliable front-end initialization of relative displacements, and accurate back-end optimization using a continuous-time trajectory model. The commonality between the proposed modules is given by the fact that all three of them exploit motion priors that are related to the inherent non-holonomic characteristics of passenger vehicle motion. In contrast to prior related art, the proposed modules furthermore excel in terms of bypassing partial unobservabilities in the transformation variables that commonly occur for Ackermann-motion. As a further contribution, the modules are built into a novel surround-view camera SLAM system that specifically targets deployment on Ackermann vehicles operating in urban environments. All modules are studied in the context of in-depth ablation studies, and the practical validity of the entire framework is supported by a successful application to challenging, large-scale publicly available online datasets. Note that upon acceptance, the entire framework is scheduled for open-source release as part of an extension of the OpenGV library.

Figures (11)

• Figure 1: Example of a vehicle-mounted surround-view camera system as studied in this paper.
• Figure 2: Notation of frames, landmarks, measurements, and extrinsic geometric transformation variables used in this work.
• Figure 3: The interpretation plane of a line $\mathcal{L}_k$ as defined by the normalized coordinates $\pi_{c_l}^{-1}(\mathbf{u}^{c_l}_{kij})$ of line measurements on a great circle.
• Figure 4: Graphical models of the different methods: (a) CBA; (b) CBARt; (c) CBASpRv; (d) SSBARv; (e) FSBA.
• Figure 5: Surround-View SLAM system overview, showing all the steps performed by the tracking and back-end optimization threads.