In this project, the goal is to write a software pipeline to identify the lane boundaries in a video.
You can find all the detail of advanced lane line detection pipeline on pipeline.ipynb. (first try)
The updated version pipeline-ColorFilterHeavy.ipynb is available. All processes are the same except for color thresholding part. I updated color thresholding logic considering additional RGB channel.
The algorithm consists of:
- Camera Calibration
- Thresholding
- Color thresholding
- Gradient thresholding
- Perspective Transform (make bird eye view)
- Lane Finding
- Inverse Perspective Transform
- Draw result!
To make video, follow generate_video.ipynb or generate_video-ColorFilterHeavy.ipynb
From project_video.mp4, Here is my first try!
The full high resolution video can be found in here: output_video.mp4
And this is my second try! (it used additional RGB channel)
The full high resolution video can be found in here: output_video_try2.mp4
After this udacity project, I thought there are so many things to be solved.
If you look at the pipeline, the algorithm seems to work well. But a lot of parameters need to be tuned over time. and I felt to use the only single algorithm for lane line detection is very dangerous. To use this kind of algorithm in real vehicle, many backup algorithms should be prepared so that roboustness can be secured.
It took about 3 minutes (5 minutes for heavy version) to process a 50-second(1261 frames) video using this framework. This means when the vehicle get where the line is, it could be too late to control with this line information. I'm concerned that this framwork can be used in a real vehicle.
This framework uses only front camera image. However, If movement of the vehicle is predictable through steering angle information, it might be helpful to improve this algorithm. By making lane line model, the tracking algorithm (such as Kalman filter or Particle filter) helps the algorithm to be more robust and accurate.