Stereo Augmented Detection of Lane Marking Boundaries

This paper presents a stereo filtering approach to augment the detection of lane markings in the Bird's Eye View image before the Inverse Perspective Mapping is applied with an existing lane detection algorithm. It also completely reformulates the semi-supervised learning process used so that it requires much less labour and is much more effective than the previously used method, allowing the approach to be extended to much larger and more complicated images. These improvements are then tested against an open and challenging KITTI data set to demonstrate the improved results. This paper then provides an analysis of the lane marking hypotheses filtered by our augmentation to illustrate that the candidates removed by our filter can produce highly confident erroneous results if not removed before Inverse Perspective Mapping due to their visual similarity to lane markings in the grayscale Bird's Eye View image.