Leveraging Vision-Based Structure-from-Motion for Robust Integrated Land Vehicle Positioning Systems in Challenging GNSS Environments

The global navigation satellite system (GNSS) is often considered as the solution of choice for navigating in outdoor environments. However, in the specific care of urban environments, the information delivered by a GNSS receiver is not very precise and may even be unavailable. For this reason, in most cases, the positioning task is realized by coupling of a satellite geo-referencing system with sensors such as odometers, inertial units, etc. To reduce the cost of such navigation solution for land vehicles, the three-dimensional reduced inertial sensor system (3D-RISS) algorithm has been developed for integration with GNSS to provide integrated navigation solution for land vehicles. The inertial sensors are accurate in the short term but drift over time. A solution is then to use other exteroceptive sensors such as cameras and Light Detection and Ranging (LiDAR). The estimation of camera's egomotion over time is known as visual odometry (VO). Although pose estimation can be achieved by evaluating the best estimate for the essential matrix using Nister's 5-point algorithm, iterative estimation from feature correspondences always causes errors that will accumulate over time and lead to drifts that even robust pose-graph optimization algorithms cannot avoid. To remedy this problem, an optimization process such as bundle adjustment is generally applied to optimize the estimation of camera poses and 3D-landmark parameters simultaneously. To make use of this optimization method, we employ structure-from-motion (SfM) techniques by exploiting the projective geometry between 3D-landmarks and their respective projections into 2D imagery. The integration of structure-from-motion algorithms for a single camera with other positioning sensors have been rarely studied, therefore the ultimate goal of this paper is to develop and present results for an integrated SfM/3D-RISS/GNSS system. The resulting solution exhibits improved positioning accuracy over the conventional 2D-2D motion estimation method under certain conditions. Experimental verification of the proposed solution is illustrated through two real road trajectories that were performed on two different land vehicles. A description is provided to show the performance of such integration in details during GNSS outages that last up to 10 minutes.