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 used a gauss kernel with a size 5.

The canny_low_threshold = 50 canny_high_threshold = 150 are used for edge detection.

After the edge detection, I used a polygon with four vertices. The bottom two vertices always have height of the image as y coordinates. The top two vertices have the same y coordinates. Instead of hardcoding the vertices, I use the factors between 0 and 1 relative to image width and height. This setting works well on the challenge video.

These parameters are used for Hough line detection. rho=2, theta=math.pi / 180, hough_thres=15, min_line_len=20, max_line_gap=10

For each line segment found by Hough transform, I calculate the line parameters assuming the lines are defined with Hesse normal form with rho and theta as parameters.

• Calculate rho and theta for each line segment. A very small float nummber epsilon is added for each division to avoid numerical error.
• Keep only the lines with absolute value of theta between min_theta=40 and max_theta=60.
• Put the lines into two groups base on its sign. Assuming that the left lane has a negative sign of theta, the right lane has a positve one.
• For each line that still remains, calculate the intersection points with the top and bottom edges of the mask polygon.
• For each line group, I use a median filter to pick the intersection point for each edge. And take the two points as the final line. I use the median filter to filter out some potential noise line segments, which have a large margin to the center.

With these methods, the pipeline works well on the challege video as well.

There are some shortcommings that I noticed:

• The pipeline is sensitive to the parameters, one have to tune the parameters with caution.
• If the images shift, the ROI will fail to cover the lanes or some noise will be there, and the pipeline will fail to detect lanes.
• In the challege video, I noticed that there are some frames with low contrast, and the pileline dese not work very well at these frames for detecting yellow lines.

In my opition, the possible improvements would be

• We may use the color (yellow and white) to filter image.
• Make the ROI somehow adaptive.
• Enhance the constrask, if, say, no lines are detected.
• Interpolate the lines across frames.