Sam's space

At the self-driving car engineer “nanodegree” on Udacity, our first project was to write a pipeline that can be used to find lane lines on an input video stream.

Processing a video sounds daunting but it’s really just a series of images that we have to deal with. And since we are given the right tools to manipuate those images, it only takes some patience and tinkering (and perhaps a sleepless night). The process is documented in a Jupyter notebook, and it included roughly the following steps:

• creating a grayscale version of the image with only one color channel:

• applying a Gaussian blur - this is done to reduce noise (Wikipedia article):

• using Canny edge detection (Wikipedia article):

• applying a Hough transform to draw lines. This uses the custom draw_lines() function, which, given a set of line segments by the Hough transformation, tries to draw the two lane lines on the image (extrapolated lines). It uses slope validation, then decides whether a line segment is part of the left or right line, averages the coordinates of the segments to have a single line, and uses a cache to fall back on if there is not enough data on the current image.

• applying a mask on the image (we only want to draw lines on the road):

• putting the lines on the original image:

This is how it looks on two simpler video inputs:

And on a more difficult one: