Lane Detector V1

Inspired after watching a video of Tesla's autopilot and my interest in automating lane-keeping in driving simulators, this rudimentary python script takes real life dash cam images and passes through a processing pipeline like one that might be found in a driver-assitance suite of most modern cars.

How it works:

Canny Operator and Thresholding

After applying a light gaussian blur to drop unnecessary details like different colors or cracks in the road surface, the canny edge detector is used to highlight edges.

Region of Interest

For a lane keeping application, the lane line should roughly be in the same area, which makes it possible to mask out unnecessary areas to only a skinny trapezoidal area in front.

From here, OpenCV's Hough-Line transform is used to determine a series of small lines. These can be separated into two main groups, those that are positive in slope, and those that are negative, indicating the left and right lane lines. These are then sorted into their respective groups, and two slopes are generated by averaging out the two containers of lines.

Finally, a translucent polygon is drawn bounded by the two lines using OpenCV and overlayed onto the original video.

This is really only for the benefit of humans, though, as the machine would only need the data behind the two opposing lines to stay in a given lane. Nevertheless, it is cool to see it come to life!

The output video:

26 seconds.mp4

Future Improvements:

My primary reservations with calling this "good enough" is its performance on curves. As shown in the video, the detection is way too erratic when the road bends.

Part of this may be due to the fact that the lane lines are quite far apart, and when on a curve, the inside of the curve is inherently shorter than the outside, meaning there is less data for a line to be found.

At around 0:30 in the video, for example, the right lane appears to be way wider than it is supposed to be, not surprising due to the lack of a lane line at the bottom right of the frame. This may be improved with using regression to fill in datapoints not present by using lane lines present in the upper right part of the trapezoid, thus predicting the probable lane line as a backup.

Lane lines at 0:30 as seen by the script. There is an alarming lack of datapoints at the bottom right of the frame, the inside of the curve.

As a version 2, I would like to use a perspective transform to allow for curves to be better identified and masked, as well as re-writing the codebase in a clean, object-oriented style for better maintainability and scalability for my plans of using it as a co-pilot of sorts for driving simulators.

Google Sites

Report abuse