Obstacle Detection for Self-Driving Cars

The objective of the project was to use stereo (depth) cameras to efficiently detect the free on-road space in front of a vehicle faster than real-time. Using the property that the horizon would always be horizontal in the images observed, we used a column-based representation (similar to stixels) to represent obstacles which resulted in a great reduction in computational complexity. OpenCV was used extensively to build a disparity map and remove road and sky using erode-dilate/watershed algorithm. We averaged the frames to reduce random noise in the image, and then calculated columnar occupancy.

A report summarizing the project can be found here: [Report]