Project overview
Finding road lane markings from a dashboard camera, solved twice: once with classical computer vision, and once with a trained encoder-decoder network.
Two approaches to the same problem, kept side by side on purpose. The first is the classical pipeline: greyscale, blur, Canny edges, a region mask over the part of the frame where the road actually is, then a Hough transform to pull straight lines out of the edges.
The second is learned. A fully convolutional encoder-decoder takes a 160 by 80 frame and emits a single-channel mask of where the lane is, so it handles curves and worn paint that the straight-line version gives up on. Having both makes the trade-off concrete: the classical one needs no data and breaks on shade, the trained one holds up and needs a dataset.