Detecting lane lines is one of the vital parts of self driving cars especially for steering. In this project, i developed a pipeline to process a video file to find lane lines and save an annotated version of the video file.
See it in action:
- https://www.youtube.com/watch?v=3p0NY46IrX4
- https://www.youtube.com/watch?v=nUdMYyYEHX8
- https://www.youtube.com/watch?v=G77kZUnWgcc
- Python 3
- NumPy
- OpenCV
- Matplotlib
- MoviePy
My pipeline consisted of 5 steps.
-
First, image is converted to greyscale
-
Gaussian blur is applied to suppress noise in the image.
-
Then, Canny edge detection is applied to find edges in the image.
-
Region of interest, described as a rectangular area is cropped not to deal with irrelevant structures in the other parts of the image.
-
Finally Hough transformation is applied and lines in the image are detected with start and end points.
I filtered out detected lines by slope. Lines that have a slope in range of [-0.5, 0.5] are ignored.
I modified the draw_lines() method in a way that line (showed in image above) is extrapolated to pass from bottom of the image and the top of the region of interest. Then top and bottom points of extrapolated lines are averaged to find only one top and one bottom point for left and right line. Then a single line is drawn between averaged points.
Also moving window average method is implemented to smooth line drawing. Drawn lines are pushed into history and then last n (window_size) frames are averaged.
One potential shortcoming would be if the lane line color and the color of the road were somehow close (in shadow or bad weather conditions), then Canny edge detector could fail to detect edges. Because, edges are detected at points which has high color gradients.
Alos, my apporach fits first order lines to the averaged points, this make impossible to detect/draw lane lines which has curvature.
White and yellow colors can be filtered to extract lane lines but not shadows, cracs and other irrelevant stuff on the road. I extrapolated every line and then averaged edge points. For an improvement to my averaging approach, standart deviation of left and right line points can be computed and then outlier points can be detected and filtered.