Lane Detection Using Computer Vision

About

This project implements Lane Detection and Advanced Lane Detection projects as required by the Udacity's Self Driving Car Nano-Degree program.

Lane Detection is done using the Python implementation of the Open Source Computer Vision Library (OpenCV).

How it Looks

A sample frame from one of the videos overlaid with markers is

Structure of the Project

The implementation is split into four files:

• LaneDetection.py: Entry point for the implentation. Read caveat first.
• LaneDetectionUtils.py: Image processing helper functions used for lane detection
• calibrateCamera.py: Routine for compensating for camera distortion before processing images and video.
• configuration.py: Configures the steps required to execute the lane detection pipeline.

Necessary input data is provided in:

• Camera_Calibration_InputImages: Set of images taken from the camera which is used in the vehicle. These images are distorted (e.g., have spherical distortion component). A standard checker-board pattern is used for computing camera calibration coefficients.
• testImages: Captured from the assignement videos provided as part of the assignments. The images are not annotated with lanes markers.
• InputVideos: A set of three evaluation videos, with increasing levels of difficultly.

Folders which may be used to hold the result of the lane-detection implementation are appropriately named.

Running the Lane Detection Application

Execute run.sh included in the project. Tested on *nix, kernel 4.x, Python 3.5.2

Caveat

Multiprocessing library has been used to speed up image and video processing. However, it was noticed that calling CalibrateCamera() routine from within the LaneDetection.py caused threaded workers to hang on a futex. See comments in lines 348-351 in LaneDetection.py. Corrections welcome.

Exploring Options

• It is possible to configure the sequence of steps executed for detecting lanes by setting or modifying options from configuration.py. Normally, no other changes are required. The changes can be made at runtime -- simply modify your choice of option, and save the configuration.py. The effect of all changes are visible after the current pool of threads returns.

• In order to evaluate the effect of the choice made, the LaneDetection continuously cycles through the images in the testImages folder, in case doLaneDetection() option is uncommented in.

• Set the number of threads allowed by tweaking the option nProcesses from the configuration.py file.

Pending Improvements

• The current pipeline works well for "project_video", and is acceptable for "challenge_video", but is poor for "harder_challenge_video". Suggestions welcome.
• The logic can be reorganized for readability.

Credits

• Udacity: Lecturers, and mentors;
• Internet: for examples and samples.

Disclaimer

Some of the ideas are borrowed and adapted from other people's work.