Page 58 - Read Online

Page 2 of 20 Huang et al. Complex Eng Syst 2023;3:2 I http://dx.doi.org/10.20517/ces.2022.43

Fusion localization methods based on HD maps have been a significant research hotspot in recent years.

For HD maps, some computer-aided generation methods have emerged [2,3] . However, lane lines obtained
usingthesemethodsareaseriesof2-dimensionalpointsets, whichoccupylargestoragespaceanddonotcarry
elevation information. Some researchers store road features in point clouds and use point cloud registration
methods to determine vehicle positions [4,5] . However, point cloud formats have disadvantages such as high
coupling, difficulty in maintaining, and unfavorable object classification. For localization, some researchers
reduce localization errors by matching road surface features, e.g. manhole covers [6–12] . However, the visibility
of road features is easily affected by illumination, which makes the matching performance differ greatly at
different times and results in unstable localization.

In this paper, the main work focuses on two aspects: First, a computer-aided generation method for HD
maps is proposed. Currently, most papers consider the lane lines are in 2D plane when the lane line fitting
is implemented. These methods are almost unusable in scenarios such as overpasses and culverts [13–16] . To
broaden the use of HD maps, it is necessary to develop 3D fitting lane lines. Second, an accurate multisensor
fusion localization method using generated HD map and existing odometry is proposed. It is worth noting
that cumulative errors will occur if the localization method is only based on odometry. Thus, a positional
constraint that has no connection with error is required to correct the estimated position. The contributions
of this paper are summarized as follows:
1. Weproposeamethodbasedonaniterativeapproximationtogeneratethe3Dcurveoflanelines. Thespatial
parameterized curve fitted by the proposed algorithm, which is global continuous, has broader applica-
1
bility than the 2D curve equation. This method not only effectively reduces the number of parameters of
the spline curve but also ensures the accuracy of the curve.
2. We separate the lanes and store them in a particular HD map format instead of holding them as semantic
information in a point cloud. For the HD maps nonuniform sampling point problem, a method based on
numerical integration is proposed to achieve uniform sampling over the arc length.
3. We propose a method to associate the elements in the HD map and the other elements in the perception
results.In this paper, the basic elements of the HD map and the complete feature associations are formu-
lated with their respective similarity evaluation metrics, considering the matching time, similarity and local
structure consistency.
4. We transform the localization problem of fusing HD maps into a graph optimization problem. Based on
the HD map and perceptual image feature association results, a lateral constraint is applied to the odometry
localization results, and accurate, low-cost localization results are obtained.

2. RELATED WORKS
2.1. Generation of lane curve equations
Chen et al. [14] demonstrated that a cubic Hermite spline (CHS) can describe line segments, arc curves, and
clothoids simultaneously and is a good choice for fitting lane lines. A CHS has at least continuity, which is
1
more accurate in describing lane curves than a traditional segmented linear fold representation. Its uniform
form allows fitting any lane curves parametrically using a sequence of feature points. Jo et al. [15] proposed a
B-spline fitting method based on the optimal smoothing technique. Zhang et al. [16] proposed a lane line fitting
method that considered a vehicle model to generate globally continuous lane lines that match the driving
1
trajectory. Gwon et al. [17] proposed a segmented polynomial fitting method with sequential approximation,
which outperformed -spline and clothoid curves in terms of computational efficiency and modifiability.

2.2. Existing HD map formats
ThereisnounifiedstandardforHDmapformats, andvariousinstitutionsandcompaniesusedifferentformats.
The OpenDRIVE standard, developed by the Association for Standards in Automation and Measurement Sys-