Feature-based Lane Detection

The feature-based methods detect the lane lines in the road images by using some low-level features, such as painted lines [17]-[19], lane line edges [7][22][23], texture [8] and colors [9][20][21]. The advantage of the feature-based methods is that the features are extracted easily. Nevertheless, they highly demand well-painted lane lines or strong lane line edges in road images; otherwise this method may fail. Furthermore, the performance of this method may easily suffer from occlusion or noise. In the following, we review some representative works of feature-based lane line detection.

Broggi and Bertè [17][18] develop an approach applying the IPM (Inverse Perspective Mapping) [42] algorithm on the road image. The IPM algorithm is a mathematical technique whereby a coordinate system may be transformed from one perspective to another, and it can remove the perspective effect from the acquired image. The perspective effect means that the lane lines width change according to

8

their distance from the camera. In this case, in order to remove the perspective effect, the a-priori knowledge exploited by the IPM transform is the assumption of a flat road in front of the vehicle. The IPM algorithm maps the acquired image, as shown in Figure 2-2(a), into a new 2-D domain array in which the information content is homogeneously distributed among all pixels and the resulting image represents a top view of the road region in front of the vehicle, as if it were observed from the top, as shown in Figure 2-2(b).

Figure 2-2 : An example of Inverse Perspective Mapping. (a) Original image. (b) Inverse perspective mapped image which seems to be observed from the top.

Here, the assumption used in the definition of a “lane line” is that a lane line is represented by a quasi-vertical bright line of constant width surrounded by a dark region (the road). Hence, the pixels belonging to a lane line have higher intensity values than their left and right neighbors. Thus, the lane line detection is reduced to the determination of horizontal black-white-black transitions. Based on a geometrical transform and on a fast morphological processing, the system is capable of detecting the lane lines. Figure 2-3 shows the results of lane line detection through the IPM algorithm.

9

Figure 2-3 : Lane line detection through the removal of the perspective effect in three different conditions: straight road with shadows, curved road with shadows, junction.

(a) Input image. (b) Mapped image (3D to 2D) of (a). (c) Result of the line-wise detection of black-white-black transitions in the horizontal direction. (d) Remapped image (2D to 3D) where the grey areas represent the portion of the image shown in (c).

(e) Superimposition of (d) onto a brighter version of the original image (a). [18]

Bertozzi and Broggi [19] propose the GOLD (Generic Obstacle and Lane Detection) system utilizing a stereo vision-based hardware and software architecture, which aims at improving road safety of moving vehicles. The GOLD system removes the perspective effect also by IPM algorithm which maps the region ahead of the vehicle into the top view. In GOLD, lane lines after the IPM transform are modeled as quasi-vertical constant width lines, brighter than their surrounding region. Based on a line-wise determination of horizontal black-white-black transitions, the pixels that have higher intensity value than their horizontal neighbors at a given distance are

10

detected. However, in this work, not only the lane line detection is implemented but also the obstacle detection.

He et al. [20] propose a color-based vision system to determine the road parameters and detect lane lines from urban traffic scenes. Based on the projective transformation, edge detection, binarization and their pre-defined curvature models, as shown in Figure 2-4, this system estimates three candidate boundaries to extract the road region. The result of boundary estimation module is combined with the color information of the capture image to get the road area image. Finally, they utilize this road area image and three candidate boundaries to determine the real road boundaries and to acquire the parameters of road. Cheng et al. [21] apply the color of road and lane lines for the image segmentation, and then utilize the size, shape and motion characteristics to determine whether a region belongs to a vehicle or a lane line for false lane line region elimination. Huang and Pan [22] develop a method to detect the lane lines and the road edges of structured and unstructured roads, respectively. In structured road detection, utilizing the vertical Sobel mask and color characteristic at first to detect the points of lane lines. If the number of points is greater than a predefined threshold, the slope filtering is applied to refine the detection results. And the least square approximation is employed to represent the lane lines. The unstructured road detection is performed while the number of detected points is not enough. They extract the road surface with the predefined sets of sampled blocks, and obtain the edge points from the vertical intervals. For the accuracy of detection, the new sampled blocks are updated by the random sampling points from segmented regions.

11

Figure 2-4 : The flowchart of He et al.[20]. (Red rectangle is represented as the curvature model they defined.)

Tsai et al. [23] propose a lane line detection algorithm using the concept of directional random walks based on Markov process. Two major components are included in this method to decide the correct locations of all lane lines: (1) lane segmentation and (2) edge linking. They first define proper structure elements to extract different lane line features from input frames using a novel morphology-based approach. Then, they utilize a novel linking technique to link all “desired” lane line features for lane lines detection. The technique considers the linking process as a directional random walk which constructs a Markov probability matrix for measuring the direction relationships between lane segments. Then, from the matrix of transition probability, the correct locations of all lane lines can be decided and found in videos.

Yim and Oh [24] develop a three-feature based automatic lane line detection algorithm

12

(TFALDA). It is intended for automatic extraction of the lane lines without the priori information or manual initialization under different road environments. The lane lines are recognized based on similarity match in a three dimensional (3D) space consisting of the starting position, direction, and gray-level value of a lane line as features, as shown in Figure 2-5.

Figure 2-5 : The lane line candidate vectors mapped into the 3D feature space.