A robust lane detection method based on hyperbolic model 9163
lines in near field part. In Sect. 4, we describe the process
of edge detection and weighted hyperbolic model matching.
In Sect. 5, the experimental results are given and analyzed.
Finally, the inference is given in Sect. 6.
2 Related work
At present, the popular traditional lane detection algorithms
are divided into two steps, first is to extract lane features
[color (Chiu and Lin 2005; Mammeri et al. 2016), edge
(Lan et al. 2009;Anetal.2013;Tuetal.2013;Wuetal.
2014; Niu et al. 2016), etc.] directly, and then fit the lane
with default models (straight line model (Saudi et al. 2008),
parabola (McCall and Trivedi 2006; Zhao et al. 2012;Yu
et al. 2017), polynomial, spline (Wang et al. 2000, 2004;
Hsiao et al. 2009; Wang and An 2010), etc.). Since the road
environment is usually very complex, the algorithms based
on feature extraction directly are affected by noise from non-
lane road, which means the result is unstable. The difficulty
of the latter is whether the default model matches the lane
and whether the parameters of the model can be acquired
correctly.
In the preprocessing stage, it is necessary to denoise (Liu
et al. 2017) and extract lane features in frames such as color,
gradient, edge, lane width and vanishing point in a complex
road environment. Chiu and Lin (2005) propose a lane mark
detection method based on color segmentation and obtain
thresholds by using gray-level histogram. Mammeri et al.
(2016) apply the HSV space to extract the yellow and white
features of the lanes for lane detection. But the detection per-
formance is affected by night environment, different lighting
conditions or passing cars. In a complete sequence of images,
the feature information of the target or the background may
change in different situations, such as illumination variation,
occlusion and scale variation (Liu et al. 2018). Since most
color models are sensitive to light, it is usually necessary
to use edge detection to extract lane features. The common
edge detection methods include Sobel (Lan et al. 2009;Tu
et al. 2013; Ozgunalp et al. 2017; Song et al. 2017), Canny
(Wang et al. 2000;Wuetal.2014; Andrade et al. 2017),
steerable filter (McCall and Trivedi 2006), etc. Niu et al
(2016) use Canny operator to obtain edge image and then
use DBSCAN to cluster the small line segments obtained by
modified Hough transform to get straight line equations of
the left and right lanes. This method performs well in most
cases, but the performance decreases obviously in the case
of large curvature. Andrade et al. (2017) use oriented gradi-
ent filter to remove unnecessary information and introduce
the yaw angle of the vehicle to correct the filtering direction.
Vanishing point (Ozgunalp et al. 2017; Piao and Shin 2017;
Tian et al. 2018) is used to control ROI region and multi-
lane detection parameters. Lee et al. (2018)use
-ROI to
constrain the lane line in perspective images.
The representation of the default model is achieved from
simple linear models to complex spline models. Saudi et al.
(2008) apply random Hough transform to detect straight lines
in the literature. This method has good performance on the
straight lanes, but cannot detect the lane with curve. McCall
and Trivedi (2006) propose a simple parabolic model, which
includes lane location, angle and curvature. The authors (Yu
et al. 2017) get the control points by getting the straight line
model first, and use the control points to fit the curve in the
parabola model, and then apply RANSAC method to opti-
mize the parameters. Although the parabolic model fits the
geometric features of lanes, the connection part between the
straight line and the curve is usually bad. To solve this prob-
lem, a parabolic model of second order is proposed (Zhao
et al. 2012). However, with the order of the model increas-
ing, it is more sensitive to noise. Wang et al. (2004) propose
a new spline-based lane model, which uses cubic B-splines
to fit the midline on both sides of the lane. In their method,
both sides of the driveway are assumed to be parallel. A lane
model based on CabMull-ROM spline is proposed (Wang
et al. 2000). It is also called Overhauser spline, that is, local
interpolating splines. Vanishing point is used to calculate B-
spline control points (Hsiao et al. 2009). S Jung et al. (Jung
et al. 2015) accumulate pixels along the time axis on the
scanning line in the sequence images and then use a cubic
polynomial and least-square method to fit the lane line.
3 Straight line combination extraction
The main purpose of this part is to achieve primary local-
ization of lanes by extracting the optimal line combination
that meet the feature, and provide parameters for subsequent
curve fitting. After the internal and external parameters of the
camera are calibrated, we can get the vanishing point of the
image. The ROI area of near field is selected as the bottom
3/4 part of the area below the vanishing point. The process
is mainly divided into three steps: line extraction based on
LSD in near field, inverse perspective transformation of lines,
clustering and selection of straight line segments.
As shown in Fig. 3, the algorithm in near field starts by
setting the ROI of the grayscale image. Then, we detect line
segments by LSD, and simultaneously the perspective and
inverse perspective mapping matrices are calculated. After
we project the segments to the ground plane with the matrix,
a histogram is applied to obtain the lane direction. DBSCAN
is used for clustering the segments, and we proposed a fusion
and extraction method to process the clusters and get opti-
mal lines combination of results. At last, the combination is
reprojected to the original image.
123