Research Project P:
A Machine Learning Approach to Lane Following

Research Project Goal

Lane-following is an important behavior for autonomous/robot driving. Often, a camera is used to find lane markers and keep the vehicle centered between them. However, this approach has some problems. First, 3D is difficult to achieve with any camera system. Thus, vision-based detection cannot benefit from 3D geometry data which is likely to provide a useful prior for the location of lane markers -- especially in areas where lane markers are worn, non-standard, or occluded. Second, vision can fail in poor lighting conditions such as early morning, at dusk, and of course at night. But all of these situations must be handled in an autonomous/robot car.

During the 2005 Grand Challenge, Stanford's winning entry utilized a fusion of vision and laser sensors to address similar issues for desert driving. This project explores how we might do the same for the new 2007 Urban Challenge. Specifically, we will use laser-based reflectivity maps (which we believe to be lighting invariant) to compensate for lighting irregularities in the visual images. Further, we will examine regions where vision-based lane tracking fails and use the laser data to assist in finding lane markers. Ultimately, these are examples of "self-supervised" machine learning, where one sensor trains another to improve its performance.

Research Project Scope

If this project is successful, it stands a substantial chance in being used on Stanford's 2007 Urban Challenge robot. Because this project is of ongoing interest to the Stanford Racing Team, the project group can anticipate significant help and support from the team including an extensive base of existing software. While this may seem to make the project easier, it actually raises the bar for this project group, since they will need to improve on existing quality research. The group must evaluate their work on real-world data sets.


  • Acquire data and software from Stanford Racing.
  • Find 10 examples of vision-based lane following failures (ie: training set).
  • Examine laser reflectivity data during these 10 failures.
  • Propose and implement fixes for these 10 failures using the laser data.
  • Re-evaluate laser/vision fusion versus vision alone on new data set (ie: testing set).

Research Project Status

student names here

Point of Contact

David Stavens,

Midterm Report

not yet submited

Final Report

not yet submitted

Course overview
Time and location
Course materials
Research Project