This is the term 1 final project from Udacity's Self-Driving Car Nanodegree program. In this project I've developed two pipelines, one for detecting lane lines, and the other for identifying and tracking a vehicle.
- Correct the distortion introduced by the camera (
camera_calfolder contains images used to undistorted the image) - Do a perspective transform on the image, so that the image represents the actual (scaled) distance between the lane lines
- Use color and gradient filtering to identify the portion of the image containing the lane lines
- Divide the image into different vertical segments and identifying points on the lane lines on left and right sides of these segments
- Fit a polynomial on left side points and another on right side points to get two continuous lines
- Reverse the perspective transform and use these two lines to color the lane
- Train a classifier to identify car vs non-car images. I've trained my classifier in 64*64 images from GTI vehicle image database and KITTI vision benchmark suite. The trained model is saved in
model.pklfile, so you don't need to re-train the model - Use the HOG features of the image to train the model
- Use sliding windows of different sizes to take out a rectangular portion of the image, resize it to 64*64 and use the trained model to predict if that portion contains a car or not
- Merge all the windows containing cars
Run main.py, it will read the video stream from project_video.mp4 and write it to processed_project_video.mp4. Note that it takes quite a while to process the video stream (25 minutes in my machine). If you want to use another video, make sure the size is 1280*720.
If you want to tweak the code, it's better to start with a single image. Comment the last four lines of main.py and uncomment the previous line. Also change the cutoff to zero of identify_windows_with_car() method in vehicle_detection.py. Running main.py afterwords will read a single image from the test_images directory and show the output.