This project is based on Udacity's Self Driving car Nanodegree
My pipeline consisted of 5 steps:
-
Convert the image to grayscale and apply a Gaussian blur. For the parameters, I used
kernel_size = (5, 5).
-
Apply Canny edge detection. In line with the recommended 3:1 ratio of thresholds, I used
low_threshold = 40andhigh_threshold = 120.
-
Apply a mask to the image outputted by Canny image detection. I applied a simple mask of trapezoid shape.
-
Apply the Hough Transform to the masked image from step 3.
-
Process and filter the lines found by the Hough Transform.
Steps 4 and 5 were performed in an iterative fashion. The Hough Transform parameters were initialized. Since the output of the Hough Transform is a potential candidates named lines, I extended these lines to the top and bottom of the masked image region. The lines were also sorted as per their length. I created a virtual zone for acceptable x coordinates of the endpoints of the line, and I removed entries in the lines list whose endpoints fell outside these intervals. I declared the longest line in lines to be one of the two lane lines, and for the other I took the longest remaining line whose slope was of the opposite sign as that of lines[0].
If this process did not yield two lines, then I altered the parameters for hough transform and repeated steps 4 and 5.
This is a simple lane detection pipeline, and as such it has a number of potential shortcomings.
- First, since we are fitting lines (not curves) to the lane lines, this pipeline may yield poor results on urban roads with sharp turns.
- Second, it may struggle when there are signs in the road, such as arrows etc.
- This pipeline may not be robust to different lighting.
To more accurately detect the lane lines, it would be beneficial to fit a Clothoid curve curve to the lanes, instead of fitting a line.