Comparative Performance Evaluation

The proposed approach was compared with the well-known methods shown in Table 2-7 [13][15]. Wang and Tsai [13] utilized a hexagon as the calibration target. However, the hexagon is not available under the moving camera, and needs to be pre-determined in the field of view. Conversely, calibration targets applied in other approaches are objects appearing in general traffic scenes, so require no additional effort on manual setting of the calibration target. The camera angle calibration in the range estimation depends only on the tilt α and the swing angle ψ, so only the access to these two angles were compared. Liang et al. [15]

assumed that the vanishing point would be in the center of the image, and accordingly estimated an approximate tilt α. Liang et al’s hypothesis is valid only in the conditions that the location of the camera is in the middle of the driving lane and the lane markings are straight lines. However, even when vehicles are driving on an ideal straight lane, it is still not easy to keep them stably in the center of lanes. Figure 2-12(a) and (b) are two cases of comparisons between Liang et al’s and our approach to estimate the tilt angle. Liang et al. [15]

proposed extending the lane markings to search for the vanishing line Vp (uv, vv) and estimating α by Vp. In Fig. 2-12, the convergent point of the u-axis and v-axis is O, the center of the image. L1 and L2 respectively represent the extensions of the right and left lane markings, and their convergent point is a vanishing point, Vp1. Liang et al’s approach estimated tilt angle by Vp1. P1 and P3 are the right and left contact points between the

preceding vehicle and the ground. The two points are applied to (2.39) to acquire the swing angle by our methods. The estimated α and estimated errors of camera angles are shown in Table 2-8, where case 1 and case 2 present the situation of Fig. 2-12(a) and Fig. 2-12(b) respectively. The camera setting in Fig. 2-12(a) is ψ=6° and α=3.5°, and in Fig. 2-12(b) is ψ=0°, α=2.5°. As shown in Table 2-8, the estimated results of tilt angle by Liang et al’s approach may have larger errors in these cases. That is because the camera is not at the center of the lane, the swing angle is not correctly estimated, and the lane markings are not straight.

Comparatively, in our method, the swing angle can be correctly obtained by (2.39) and then the tilt angle can also be appropriately estimated by (2.42). Therefore, in these cases, our approach can obtain more accurate results without the limitations due to some pre-determined conditions. Among the three approaches in Table 2-7, only Liang et al’s and our approach use calibration targets on the road to achieve dynamic calibration, when the moving camera causes continuously variations of tilt angle α.

Table 2-7 Comparison of Approaches

Table 2-8 A comparison in estimation results of camera angle and errors Case 1,

Approach Calibration Target Calibration angle Occasion

Wang and Tsai [13] Hexagon ψ, α Fixed camera

Liang et al. [15] Lane marking Approximation of α Moving camera Our approach Lane marking,

Vehicle

ψ, α Moving camera

(a)

(b)

Fig. 2-12. The swing angle calculated by Liang et al’s and our approaches. (a) Straight lane markings. (b) The curve of lane markings.

Chapter 3

Lane Detection

3.1 Introduction

In the driving assistance systems, traffic information can be acquired by sensors to make driving safe and easy [81][82]. For example, vision-based driving assistance systems can determine positions of lanes and obstacles preceding a host vehicle, and the detected information can serve as guidance for driving safety of vehicles [83]-[85]. In the system, the detection of lane is based on image processing techniques to search for the road edges or the lane markings [37][59] and then the lane information is applied to the detection of obstacles in determining obstacle positions [7][33][70]. However, occlusions of obstacles on lane markings may affect results of lane detection [86]. Therefore, lane detection requires not only fast executive speed to achieve real time detection, but also a solution to occlusions.

This dissertation applied geometry transformation and a method of rapid computation of lane width to predict the projective positions and widths of lanes and markings. Then, an approach named LME FSM is designed to find lane markings efficiently. A statistical search algorithm is also proposed to correctly and adaptively determine thresholds under various illumination conditions. Furthermore, a dynamic calibration algorithm is presented to update the information of a camera’s parameters and lane widths. Besides, a fuzzy logic scheme is adopted to judge the correctness of the detected lane markings and the results are applied to the selection of knots when reconstructing road geometry by B-spline. Finally, the ROI determination strategy is proposed to constrain the search region to make the detection more

robust and fast. Therefore, even though obstacles occlude parts of the lane markings, road boundaries still can be reconstructed correctly. Besides, the relative positions between lane markings and cameras can be more precisely estimated with the camera tilt obtained through dynamic calibration.

The rest of this chapter is organized as follows: Section 3.2 presents image analyses using a camera model and the approach of dynamic calibration; Section 3.3 describes the proposed approaches to lane detection, including analyses of lane features, a novel lane marking extraction method adopting a finite state machine, a strategy for determining ROI, post processing by fuzzy reasoning, the determination of road boundaries by B-spline curve fitting and overall process of lane detection. Then, the experimental results of the lane detection and analyses are shown in section 3.4.