When we drive, we use our eyes to decide where to go. The lines on the road that show us where the lanes are act as our constant reference for where to steer the vehicle. Naturally, one of the first things we would like to do in developing a self-driving car is to automatically detect lane lines using an algorithm.
In this project we detect lane lines on the road from images and videos taken by a camera on the car.
The file containing the project code is P1.ipynb, which is a Jupyter Notebook. Python and OpenCV is used to implement the algorithms. This Notebook is exported as an HTML file to P1.html for easier viewing of the results.
The requirements of the project are detailed here: project rubric
In order to detect lane lines on an image or frame of a video stream, we apply the following pipeline to the image or video frame:
- Convert image to grayscale
- Apply Gaussian Blur smoothing to reduce noise for the edge detection step
- Apply Canny Edge Detection
- Apply region of interest to mask the lines out of this region
- Apply Hough transform to detect lines
- Draw lines by following these steps:
- Separate lines into right and left based on being negative or positive
- Filter out lines with unreasonable slope values by comparing to some thresholds
- Calculate a weighted average of slopes
- Fit a line to left and right points
- Overlay left and right lines onto the original image
The performance of the lane finding algorithms may degrade in these situations:
- Low color contrast between road and lines, e.g. due to rain
- low lighting e.g. at sunset
- Curved roads
- When there is no lanes on the road, e.g. on dirt road
- When lanes are temporarily not straight, e.g. when merging with another road
- When road goes on a hill or vertical curvature
- When other cars cover the lanes
We can use these methods to make our algorithm generalize better for a wider range of circumstances:
- Non-linear fit for the curved lanes
- Use parameter filtering to make changes from one frame to another smooth
- Impose a maximum change in parameters (e.g. slope of a line) from one frame to another. e.g. a parameter cannot change more than 10% from one frame to another
This project is copyright is under MIT License.