Project P16:
Road surface type estimation for the DARPA Grand Challenge

Project Goal

Stanford is participating in the DARPA Grand Challenge, an autonomous robot desert race. So far, our robot relies on camera video and laser scanners to detect road patches that are safe for our vehicle to traverse. While determining the location of the road quite well, the algorithm does not yet include an estimation of the kind of terrain our vehicle is on. This information is not only important for selecting a safe velocity, but also potentially useful for setting controller parameters based on the known slipperiness of a terrain type.

Project Scope

Your algorithm will accept as an input a video stream and a mask of where our racing software has located potentially drivable surface. If necessary, it can further utilize other data provided by the race vehicle software, such as time of day, vehicle position/orientation or 3d structure of the terrain. Your task will be to compare this area of drivable surface to various known terrain types, including sand, rocks, mud, water, concrete and asphalt. Determine the most resembling type as well as a confidence measure. Since terrain type does not change with every video image, your algorithm does not need to run in realtime, a computation time of ~10 seconds should be acceptable.

Tasks

• We have about 10 hours of desert driving video. Select a representative set of images for the different terrain types as training data.
• Select a set of features and properties useful for distinguishing terrain classes. You may want to look into fourier coefficients, histograms of the HSV color space and others.
• Apply learning or statistical algorithms such as support vector machines or linear models to learn the category discrimination.
• Evaluate your categorizaton on testing sets different from the ones used for learning.

Project Status

Kyle Heath (heathkh at stanford),
Eric Liang (eliang at cs),
Yu-Yao Chang (yychang at stanford),
Ari Steinberg (ari.steinberg at stanford)

Point of Contact

Hendrik Dahlkamp

Midterm Report

not yet submited

Final Report

not yet submitted