dc.description.abstract | In recent years, there were lots of deaths caused by lane departure of traffic accidents. To prevent these fatalities, lane departure warning (LDW) systems have been proposed to protect drivers from the departure accidents. A standard LDW system should include two major parts: lane-mark detection and departure measurement. The lane-mark detection extracts lane marks, and then the measurement module acquires the lateral offset of the vehicle to adaptively prevent the lane departure.
In this dissertation, we present a versatile lane departure warning system based on three-dimensional (3D) visual geometry that would help drivers avoid unintended departure from their lane during land vehicle driving. A horizontal gradient map was first calculated by an edge operator combined with the properties of lateral inhibition and far-near adaptation so that the operation would be less influenced by bad weather conditions. The lane marks were then detected with the proposed conjugate Gaussian model on the non-thresholded gradient map to make the detection more stable and less influenced by shadow boundary, windshield wipers, and the partial occlusion of other vehicles. The detected lane marks were then verified by the 3D geometric constraints of the parallel lines on the road surface to avoid the wrong detection of non-lane marks; moreover, the constraints were also used to find the other-side lane mark from the extracted one if only one lane mark was extracted on low-quality road surfaces. This means that the lane can be detected even if only one lane mark is detected. Lastly, the lane offset distance was accurately calculated from the 3D geometry of the lane marks rather than roughly estimated from the two-dimensional relationship of the lane marks on images.
For efficiency, straight-line-based lane detections were generally utilized to detect straight or curved lane marks. However, using straight line to approach curved line may results in error and the error is influenced by variant situations. Here we want to know whether the straight-line-based LDW system is sufficient to work for various road situations. In this dissertation, we analyze the errors of the estimated lateral offsets which were generated by a straight lane model in three different kinds of road situations: (i) the accuracy of the estimated lane width, (ii) the horizontal curvature of the lane marks, and (iii) the vertical curvature of the lane marks.
Based on our analyses, the properties of the proposed system are special and unique and are not present in the existing systems. Based on various images of weather conditions and road surface situations, the detection rate reaches over 97.1%. This study demonstrates the effect and efficiency of the proposed system and compares it with other existing systems. Besides, according to the analysis on various geometric-shaped artificial curved lanes and the experiments on real images based on our experimental platforms, we find that most errors of estimated lateral offset are insignificant in general driving situation; on highway, the maximum errors reach 7.1 cm of horizontal curved lane and 13.7 cm of vertical curved lanes on highway. However, those maximum errors occur on the far lane mark when crossing lane. In that moment, the lateral offset is estimated by the lane mark which driver crossing. The maximum errors in general driving situation, includes the lane mark which driver crossing when changing lane, are only 3.8 cm of horizontal curved lane and only 2.2 cm of vertical curved lanes. That is, the straight-line-based lane departure warning systems are reliable to provide drivers the lane departure warning information whenever driving on straight or curved lanes. | en_US |