A novel path segmentation method for autonomous road following

Detecting roads using monocular vision is a very challenging task as the detection algorithm must be able to deal with complex real-road scenes. In this paper, we describe an algorithm for general path segmentation. There are three main technical contributions of the approach. First, a path segmentation framework is presented, which formulates road detection as a Bayesian posteriori estimation problem. Second, to obtain knowledge about the road surface and encode drivable region, a self-supervised learning algorithm is proposed. Third, three image based statistically independent measurements are developed, which are then incorporated into the Bayesian framework to find the two most dominant edges of the path. The method makes no assumptions about the structure or appearance of the road. Experimental results demonstrate that it achieves substantial improvements over the method which represents the current state of the art in unsupervised and self-supervised path segmentation research. Video clips that can be found at http://pan.baidu.com/s/1hqSq8vy show our segmentation results under five different challenging scenarios.