B$^2$F-Map: Crowd-sourced Mapping with Bayesian B-spline Fusion

TL;DR

This work presents a complete pipeline for HD map generation using production vehicles equipped only with a monocular camera, consumer-grade GNSS, and IMU, and proposes a novel Bayesian fusion framework for B-spline trajectories with differing density representation, enabling principled handling of uncertainties.

Abstract

Crowd-sourced mapping offers a scalable alternative to creating maps using traditional survey vehicles. Yet, existing methods either rely on prior high-definition (HD) maps or neglect uncertainties in the map fusion. In this work, we present a complete pipeline for HD map generation using production vehicles equipped only with a monocular camera, consumer-grade GNSS, and IMU. Our approach includes on-cloud localization using lightweight standard-definition maps, on-vehicle mapping via an extended object trajectory (EOT) Poisson multi-Bernoulli (PMB) filter with Gibbs sampling, and on-cloud multi-drive optimization and Bayesian map fusion. We represent the lane lines using B-splines, where each B-spline is parameterized by a sequence of Gaussian distributed control points, and propose a novel Bayesian fusion framework for B-spline trajectories with differing density representation, enabling principled handling of uncertainties. We evaluate our proposed approach, B$^2$F-Map, on large-scale real-world datasets collected across diverse driving conditions and demonstrate that our method is able to produce geometrically consistent lane-level maps.

Figures (5)

• Figure 1: System overview of B$^2$F-Map pipeline, including three modules: on-vehicle localization, on-vehicle mapping, and on-cloud localization and mapping. Note that traffic signs in the local maps are represented as semantic points and lane lines are represented by B-splines. With B-splines continuous over time, it is bandwidth efficient when uploading the control points to the cloud. On the cloud, after optimization to eliminate positioning errors in the estimated lane lines, the Bayesian B-spline fusion algorithm performs map fusion while maintaining the same B-spline representations.
• Figure 2: Visualization of a split/merge area. After optimization, positioning errors are reduced - lane lines observed multiple times are mostly stacked on top of each other. After fusion, redundant representation is removed.
• Figure 3: Illustration of two B-splines, completely overlapping (closely-spaced) in (a), partially overlapping in (b)-(h). Case 1-5 are presented in details in Section \ref{['sec:fusion_two_partially_overlapping']}. Note that (c) and (d) both correspond to Case 2, where in (c) the two trajectories would be merged into one whereas in (d), the trajectory in blue will be truncated into two parts. Similar for Case 3 in (e) and (f). (g) illustrates Case 4 where an interior subsequence of one B-spline overlaps with a subsequence of another. (h) illustrates the case where the overlapping area is discontinuous, e.g., when traffic islands are present.
• Figure 4: Qualitative comparisons of clustering and assignment DA using Murty's algorithm Yuxuan2024Fusion (baseline) in (a) and Gibbs sampling-based DA in (b). The blue points are the noisy lane marking edge detections and the red lines are lane lines produced by EOT tracker. The gray arrow in (a) illustrates the vehicle's travel direction. (c) shows the Google street view of the lane markings from another lane line, which are wrongly associated in baseline approach, but not in the Gibbs sampling based approach.
• Figure 5: A qualitative result on the crowd-sourced map, where the fused lane lines from B$^2$F-Map pipeline are in black, and the ones from baseline are in red. The estimated lane lines from crowd-sourced vehicles before fusion are in blue and the ground truth is in green. Note that B$^2$F-Map results in accurate estimates despite the errors in some lane lines before fusion.