This is the first lane lines project for Udacity's Self-Driving car nanodegree program. It is using relatively ad hoc methods to detect the lane lines on a road in moving video. The general ideas in use are:
The process used in this project is very basic. Roughly the following steps:
-
First the color is filtered to eliminate non-line components. This is performed by simplistic color selection for white and yellow.
-
Next, Canny filtering is applied. This involves blurring the image and thene edge detection using the Canny filter, which is a filter based on the absolute values of the Cartesian partial derivatives of the image. This results in a collection of points that identify the edges of the lines. Unfortunately, this also identifies the edges of other features like bridges and signs.
-
A Hough transformation is applied that allows us to think of the points in terms of line segments with various slopes. These slopes can be bounded as a means of rejecting aspects of the image that are not what we would expect of road lines.
-
A filter region can be applied to the visual area to further reduce artifacts that are not in the field of view that we would prefer. This leaves us with only the line areas, their edges, and their slopes.
-
We can identify these areas and use a moving average filter to perform smoothing. What I chose to do is use the average slope and the average nearest extrema for each road line as a means of generating a fixed length line through a point near the car. The reason for this is that it looks better and more consistent.
The unprocessed and resulting videos are in the repo if you are interested. Here is one of the videos posted to YouTube.