Lane Detection

There are many ways lane detection can be achieved. In early studies, Dickmanns et al.

[26]-[28] conducted 3-D road recognition by adopting horizontal and vertical mapping models, the approach of extracting features with edge elements, and recursive estimation techniques.

The results were applied to their test vehicle (VaMoRs) to function as autonomous vehicle guidance. Broggi et al. [29][30] used IPM (Inverse Perspective Mapping ) to transfer a 3-D world coordinate to a 2-D image coordinate, and detected road markings using top-view images. Kreucher and Lakshmanan [31] suggested detecting lane markings with frequency domain features that capture relevant information about edge-oriented features. The objectives of many studies on lane detection include autonomous vehicle guidance and driving assistance such as lane-departure-warning and Driver-Attention Monitoring systems. Some

assumptions in common are as follows: 1) The road is flat or follows a precise model. 2) The appearance of lane markings follows strict rules. 3) The road texture is consistent. The main difficulty in lane detection is how to recognize roads efficiently in various situations, including complex shadowing and changes in illumination [32][33]. Furthermore, the vibration of a moving camera causes changes in camera parameters and thus leads to errors in geometric transformation. To solve the problem, dynamic calibration of cameras is required to improve robustness [34]-[36].

The task of lane detection can be summarized as two main sections: 1) The acquirement of features. 2) A road model for reconstructing road geometry. In addition, to accelerate the detection and make it robust, some approaches are added such as narrowing the search region, the determination of ROI (Region of Interest), dynamic calibration for the camera and position-tracking methods using consecutive images.

The first step of detecting lanes is to extract their features. On most occasions there are lane markings on both the left and the right side of the driving lane, while sometimes only the boundaries of the road exist without any lane marking. Most parts of the lane markings are like two parallel ribbons with some variations, for example, being straight or curved, solid lines or dashed lines, and in the color of white, yellow or red. The occlusion of trees and buildings and their shadows makes it more difficult to detect positions of lanes. Also, visibility varying with illumination increases difficulty to detections [37]-[41].

In acquiring features, there are four major types of methods: pixel-based, edge detection-based, marking-based, and color-based methods. The pixel-based type is to classify pixels into certain domains and put pixels of the road boundaries in one category [42]-[44].

The edge detection-based type involves conducting edge detection in the image first. Then, find straight lines with Hough transform [45][46] or adopt an ant-based approach to start a bottom-up search for possible path of the road boundary in the image [47]-[50] or determine

search regions by road models for detection of road boundaries [1][51]-[53]. Those two methods are time-consuming, and easily cause errors when complex shadows or obstacle occlusion exist. The marking-based type is based on features of lane markings. For example, Bertozzi and Broggi [29] proposed IPM and black-white-black transitions to detect lane markings. This method may effectively deal with some situations of shadows or obstacle occlusions. However, the vibration caused by the moving vehicle may influence the extrinsic parameters of the camera, and thus arouse unexpected mapping distortions on images, which may cause errors on lane detection results. The color-based type is to utilize color information of the road in the image [44][54]-[56]. In this way, there is more information about the lane and better abilities to resist noise. However, it takes more computation time to extract color features of interest.

Since shadows of trees or other noises usually exist and some lane markings are dash lines, the detected features of road boundaries are often incomplete. Therefore, the methods of interpolation or curve fitting are needed to reconstruct the road geometry. Kreucher and Lakshmanan [31] used a deformable template shape model to detect lanes. They believed that two sides of a road respectively approximate a quadratic equation, so they established their coefficients to determine the curvature and orientation of the road. However, curve fitting cannot be done by a quadratic equation on the lanes with S-shaped turns. So Wang et al. [57]

adopted spline interpolation which can be used in various curves to connect line segments.

However, when there are vehicles in the lane occluding parts of the lane boundary, some errors may arise, because this approach found a vanishing point depending on Hough transform followed by line-length voting. Thus, vehicles on the lanes may form spurious lines which may influence the determination of the vanishing line. Furthermore, Hough transform and Canny edge detector utilized in Wang et al.’s approaches may take more computation time.

Another issue to promote lane detection efficiency and depress noise sensitivity is to set appropriate ROI. Lin et al.[58] applied the information of both lane boundaries obtained from initial detection in the first frame to the finding of ROI on Hough domain. Then ROI was adopted as search parameters of lane boundaries in the subsequent frames. The method can effectively accelerate lane detection process on a straight lane but errors may arise on road curves. Chapuis et al.’s method [59] utilized an initially determined ROI to recursively recognize a probabilistic model to conduct iterative computation, and adopted a training phase to define the best interesting zone. The initially set ROI is effective in the general roads;

however, diverse road curves may make the initial ROI too large on the farther part of the road and thus raise noise sensitivity. Therefore, an effective method is needed for adaptive determination of ROI and adjustment to changes of road curves in the image sequences, and thus ROI can be significantly narrowed to obtain more accurate and faster lane detection results.