Experimental Results

Here, we implement several related methods on the intermediate case, the lane changing case and the straight lane lines case, and then we show the experimental results. We discuss the problems occurring in the experiment. Since the intensity of a lane line is brighter than the neighboring road surface in the image, Nadra et al. [36]

utilize the top-hat transformation, which is one of the morphological operations, in the pre-processing step to extract the clear regions in the image regardless of background variations, as shown in Figure 4-2(b). For the contrast enhancement, as shown in Figure 4-2(c), they give a threshold based on the top-hat transformed image to enhance the intensity of the pixels. However, the threshold is hard to define. Another question is how to select the structure element of top-hat transformation for lane line detection. If we apply the 33 structure element and set the threshold as 6 in the top-hat transformation to test our dataset, some lane lines may be destroyed, as shown in Figure 4-3(b). Besides, due to too many noise pixels included in the enhancement image, as shown in Figure 4-2(c), the recognition rate of lane lines decreases. By the way, before pre-processing, they divide the image into left and right parts, as shown in Figure 4-2(a), as the regions of interest. This procedure makes the system only can detect the condition where the lane lines are located in the left and right parts. When doing the lane changing, the lane line crossing the middle part is not detected.

Figure 4-2 : The results of pre-processing step in [36]. (a) Original image. (b) Image after top-hat transformation. (c) Image after contrast enhancement.

51

Figure 4-3 : The result after the top-hat transformation and contrast enhancement. (a) Original image. (b) Image after top-hat transformation and contrast enhancement where the red cycle shows the destroyed part of the lane line.

Hence, for reducing the noise pixels in the textured areas, we apply an additional constraint “structure tensor” to find those line-structure pixels on the contrast-enhanced image. By observing the two eigenvalues of the structure matrix SM which is computed over a small window of size (2q+1) around each candidate effectiveness of this method in removing the word pixels in the bottom right corner of the image and the pixels of the grass or the wall. Nevertheless, the disadvantage of this method is that it takes more than 300ms, which is not suitable for the real-time requirement. For the reason of above paragraphs, we do not apply the top-hat transformation and line-structure constraint in our system.

As described in Section 3.3.2, the Hough transformation is one of the most common algorithms to detect the straight lines in an image. For various types of lane lines, there should be many lane line models used to describe them. The straight and

52

curve are the common lane line models. In this thesis, we first detect and record the lane lines from the straight, and then track them in subsequent frames. Figure 4-5 shows the result of lane line detection when a new lane line appears in the middle of two originally detected lane lines. Result of the intermediate case of driving from the straight lane lines to curve lane lines then back to straight lane lines is shown in Figure 4-6. Figure 4-7 and Figure 4-8 show the result of the lane changing case and the straight lane line case, respectively.

Figure 4-4 : The result of line-structure constraint. (a) Original image. (b) Image after top-hat transformation and contrast enhancement. (c) Image after the line-structure constraint on (b). The red cycle shows the effectiveness of noise removal.

Figure 4-5 : A new lane line appears in the middle of two originally detected lane lines.

(a) Original images. (b) Output images.

53

Figure 4-6 : The intermediate case of driving from straight lane lines to curve lane lines then back to straight lane lines. (a) Original images. (b) Output images.

Figure 4-7 : The lane changing case. (a) Original images. (b) Output images.

Figure 4-8 : The straight lane line case. (a) Original images. (b) Output images.

54

Although our algorithm applying the time-slice images to track the lane line points is feasible, there are two problems arising in our experiments. Problem 1 is that the number of times the gradient value smoothing step is performed affects the declining speed of the gradient histogram. As described in Section 3.4.3, we smooth the movement of gradient histogram and reduce the generation of the wrong peak points when the lateral moving of the vehicle occurs. Nevertheless, the more number of times of smoothing, the more quickly the gradient histogram decreases. This phenomenon makes the detection of the dashed lane lines go wrong easily. Figure 4-9 shows an example of incorrect lane line detection due to too many times of smoothing steps. We focus on the left lane line and use four times of smoothing step. Since its internal distance is farther than the right lane line, we loss its information on the first ROI (red color) at frame 458, which is just 3 frames after frame 455. Since from frame 458 to frame 464, the next left lane line segment does not pass our first ROI yet, we loss the left lane line information on the second, third, fourth, and fifth ROI in the image, respectively. Although the new left lane line segment passes our ROIs at frame 465, the left lane line in the tracking list has been destroyed since it misses for too many frames. On the other hand, as shown in Figure 4-10(a) and (b), we apply 1 time smoothing and 4 times smoothing in the gradient histogram, respectively. The gradient histogram after 1 times smoothing generates many peaks in the ROIs, which makes the result of lane line detection go wrong easily. The yellow regions show the difference between Figure 4-10(a) and (b).

55

Figure 4-9 : Problem 1 caused by the smoothing step. (a) Original images. (b) Output image where the left lane line is incorrect at the frame 465.

Figure 4-10 : The results under the different times of the smoothing step. (a) After 1 time of smoothing step, our system generates an incorrect lane line because of too many peaks in the gradient histogram. (b) After 4 times of smoothing step, the result of lane line detection is almost correct because of the proper gradient histogram.

Problem 2 is how to check whether the shape of a lane line is correct when tracking both on the straight lane lines and curve lane lines. Even though we already

56

obtain a lane line positions on each ROI in the detection step, the new point positions on each ROI tracked correctly cannot be ensured because of the noise. The noises may come from the leading vehicles (c.f. Figure 4-11 (a)), the words (c.f. Figure 4-11 (b)), and the reflection of the windshield (c.f. Figure 4-11 (c)). They cause errors in lane line detection. Furthermore, if we find out that several tracked points of a lane line are wrong, how to modify this lane line to be correct is also one of the challenges in our system.

Figure 4-11 : Problem 2 caused by the noises. (a) Noise from the leading vehicle. (b) Noise from the words. (c) Noise from the reflection of the windshield.

In the end, we also implement the method proposed by Nadra et al. [36] in our experiments. To compare with this method by the statistic analyses, three performance indices are used, i.e., missing rate, false detection rate, and accuracy. The definitions of missing rate (MR) and false detection rate (FR) are described by Eq. (21) in Section 4.2. Table 4 lists the detailed performance comparisons between [36] and our method. From the view of accuracy, the method [36] performs better than our method for the Clip#2 (lane changing case) and Clip#3 (straight lane line case) since we have not resolved the Problem 1 and Problem 2 yet. But for the Clip#1 (intermediate case), their method performs worse than our method because their method needs to delimit a region for the curve fitting method. However, the performance of their system does

57

not work well on our dataset because of two reasons. One reason is that they select the maximum voting line by the Hough transformation as the lane line in the left and right region of interest, respectively. But they do not mention about how to select the better region of interest in their system for different car images. The other reason is that they only detect the lane lines for each single frame, and do not include the tracking conception. Looking back at the performance of our system, since the challenges mentioned before still need to be conquered, there is still room for improvement. In conclusion, if we can overcome the problems arising in our experimental results, our system utilizing the time-slice images to detect and track the lane lines will become feasible and robust.

Table 4 : Performance comparisons between Nadra et al.[36] and our method.

Clips

58