Finding Lane Lines on the Road
The goals / steps of this project are the following:
- Make a pipeline that finds lane lines on the road
- Reflect on your work in a written report
- I converted into grayscale but then I found that converting into HSV is more reliable when working on two (or more) colors. I used GIMP to find color saturation for yellow.
- Softened the image by applying gaussian by the degree of 3. 5 is also suitable but sometimes it can cause lost of line because of low number of edges.
- As mentioned Canny's higher threshold should be between 2 to 3 times of lower threshold. So I defined a factor to make coding easier.
- Applied mask so I don't get unnecessary points on images.. Also draw mask lines on image so I can get better understanding of where we are looking at.
- Hough transform with threshold=35, min_line_length=5, max_line_gap=2. This is bit trial, not much science in it (at least for me :)) 6.draw_lines() function: I tried many ways to find the right solution. First idea was to find all the points according to their slopes and adding them to lists. Then getting upper left/right and lower left/right points, tried to draw lines according to those points. But upper right point was always on the left half of the image so it was impossible to draw a right lane. Then I tried np.polyfit but it didn't give me what I really wanted. So I tried easier way, found the mean of the slopes and the points for left and right lanes. Draw the lanes according to these.
Although pipeline seems to work fine with example images and videos, at challenge video while crossing the bridge, (as far as I could see) number of points increases suddenly to a large amount and lanes gets into a Star Wars laser saber fight. Applying higher gaussian kernel size to make image smoother did not work as I expected. I saw many people were good at drawing lines during passage of the bridge, so I should be seeking for another solution. And also lane lines is being drawn longer than the mask. It was getting wrong image size, before fixing it lines were starting higher than the bottom of the image. I managed to fix it but not passing the mask problem.
Challenge video can be much better than this.. And my code is bit hard to follow (even for me), I should learn to code more clearly.
Left and right lane angles narrowed down using np.arctan() so that only the lines which are between 20-60 degrees accepted. This helped crossing the bridge part. Introduced last_working_right_line and last_working_left_line variables which holds the data of the last known location of the lanes so that if we cannot draw any line because of missing points, we go back and collect the last working points. By this way (checking by math.isnan) we increased quality and stability of the lines.
Variables changed: rho=2 threshold increased to 50 min_line_length=20 max_line_gap=100
as advised.
Fix:
Removed ROI mask from video.
TODO:
Fix lane lengths: When I increase 0.6 to 0.7c(or more) (mask_y_height=int(img_shp[0]*0.6)) I got an error. Seems like I am not able to find any points when mask is increased. I may be able to fix this if I find a way to cut lines according to ROI.