الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 78 |  |  |
| | | from | 78 | | |
Abstract
With the increase in number of accidents on the world roads, many intelligent
safety systems such as the advanced driver assistance systems (ADAS) have been developed to provide assistance to the drivers in order to eliminate or decrease the human error ; hence, increase the vehicles safety. The lane detection provides information which helps locate the lane boundaries on the road in front of the ego-vehicle. Thus, it’s a crucial and fundamental component of various ADAS including the lane keeping assistance and the lane departure warning assistance systems. The lane detection task confronts several challenges in urban roads including curved lane boundaries, presence of street writings, faded lane markings and inconsistencies in illumination conditions.
In this thesis, a robust and vision-based proposed method for lane detection that demonstrates stable detection performance in various illumination conditions and driving scenarios in urban roads, is presented. The proposed method defines a region of interest (ROI) to exclude the misleading parts of the road image that may confuse the features extraction step. This also reduces the computation time of the subsequent steps. For the purpose of eliminating the perspective distortion appearing in road images, inverse perspective mapping (IPM) is employed to obtain a bird’s-eye view (BEV) of the road in front of the ego-vehicle. We propose an approach for representing the color information of the BEV image in order to enhance the extraction of the lane features of interest and reduce the outliers. This approach is based on fusing the saturation and value components of the hue saturation value (HSV) color model into a single fused channel. Afterwards, lane markings detection is carried out to produce a binary image where the white pixels represent the potential lane points. The estimated lane boundaries are then represented by parabolic models whose parameters are estimated from the potential lane points using the random Sample Consensus (RANSAC) algorithm.
Furthermore, we present a thorough evaluation of the detection performance of
proposed method using the ground-truth lane points of the Caltech dataset. This
performance evaluation provides quantitative assessment that can be used to find the optimal configuration parameters of the proposed method. The detection results show that the proposed method detects lanes in various road scenarios and driving scenes. In addition, a comparative analysis is conducted between the proposed fused channel approach and the conventional gray scale intensity one that is often utilized in the literature. The fused channel approach achieves F1 scores of 0.92, 0.68, 0.85 and 0.93 on the four video sequences of the Caltech dataset. Moreover, it outperforms the gray scale intensity approach by 16.31% and 10.42% on the two video sequences where more complex situations such as illumination variations, shadows casted by trees, passing vehicles, street writings and lane changes are present.