1.1 Motivation and Overview
Recently, traffic accidents have become one of the most serious problems in today’s world. However, the major factor which leads to road accidents is the carelessness of drivers or improper driving. Therefore, Advance Driver Assistance System (ADAS) [1] and Intelligent Transportation System (ITS) [2] are developed to improve the driving safety and reduce road accidents.
For the vision-based driving assistant systems, lane line detection plays an essential role in providing useful and effective information with respect to the relative position of the vehicle on the road. By means of this information, the driver can better understand the road circumstances and his/her driving situations for safety. So for decades, lane line detection has become a critical research field. However, in most conditions, vision-based lane line detection is simplified into a problem of finding the locations of lane lines in the input road images with or without strong prior knowledge about the lane line positions and drawing the results in the output images. Figure 1-1 shows an example of the lane line detection.
In order to analyze the lane line information from car videos, most lane line detection algorithms are based on image processing techniques to search for the lane lines. In general, the video analysis procedure comprises three major processing steps:
(1) selection of the region of interest, (2) lane line detection, and (3) lane line tracking.
Nevertheless, in most of the existing works, a fundamental problem is that the performance of video analysis may not be stable with varied environments and different weather conditions, as shown in Figure 1-2, resulting in the difficulty in lane
2
line detection. Besides, as shown in Figure 1-3, the presence of shadows, the lane line occluded by the vehicles and various markings on the road also affect the detection result. Moreover, most existing works processed the case of straight lane lines and curve lane lines, individually. Few researches discuss about the intermediate case of driving from the straight lane lines to curve lane lines, as shown in Figure 1-4, or lane changing, as shown in Figure 1-5. On the other hand, another critical issue is that image processing is always time-consuming. To cater for the requirement of real-time response, developing efficient system frameworks and video analysis algorithms is an important and inevitable task. Consequently, how to quickly and correctly locate the position of the lane lines from car videos is the core problem and issue in our work.
Figure 1-1 : An example of the lane line detection. (a) Input frame captured from the camera. (b) Output frame where the position of the lane lines is located.
Figure 1-2 : Different weather conditions. (a) Image captured under the sunny day. (b) Image captured under the cloudy day. (c) Image captured under the rainy day.
3
Figure 1-3 : Different driving environments. (a) The presence of shadows. (b) The lane line occluded by the vehicles. (c) Various markings on the road.
Figure 1-4 : An example of the intermediate case of driving from the (a) straight lane lines to (b) curve lane lines and then back to (c) straight lane lines.
Figure 1-5 : An example of lane changing from the left to right.
In this thesis, we describe our lane line detection and tracking system, then implement and compare several methods on the various cases, as shown in Figure 1-4 and Figure 1-5, and we discuss some problems arising in the course of our experiments. We mount the camera on the upper center of windshield of the vehicle for video capturing when driving. When inputting a car video, our proposed system
4
locates the vanishing point at first, and then instead of using the whole image, we only process some rows of the image as our rows of interest (ROIs). Slicing through the stack of some frames in time at a ROI generates a time slice image. With the time slice images and the vanishing point’s position, we can detect the lane lines and track them on the time slice images. In the experiment, we implement several lane line detection algorithms and test them on real car videos captured at special environment conditions, then we discuss the problems arising in the experiment.
In conclusion, the contributions of this thesis are listed as follows:
We adopt the peak finding algorithm to find out the points of the candidate lane lines, instead of giving a fixed threshold to classify the pixels in the image into non-lane-line and lane-line classes.
We propose a gradient value adjustment algorithm to overcome the sparseness problem in detecting dashed lane lines.
We propose a lane line detection and tracking algorithm using the “time slice images” which involve the lane lines’ relationship between the time and space.
Also, the time slice images can help us to detect the shape of curve lane lines without calculating the parameters of curve fitting.
1.2 Organization
The rest of this thesis is organized as follows. In Chapter 2, we survey some related works in lane line detection and tracking of car videos. Chapter 3 introduces our system to detect and track the position of lane lines. In Chapter 4, the experimental results of lane line detection and tracking are shown and discussed. At last, we conclude this thesis and describe the future work in Chapter 5.
5