Texture-based segmentation of road images

Abstract

The extraction of road boundaries is one of the basic requirements for an autonomous navigation system to guide a vehicle on a road. The existent road detection approaches have some difficulties with the roads that have neither marked lanes nor different colors between their surface and the environment. Instead of gray value or color, texture can be an important feature in such images. Taking this as motivation, the authors developed a texture-based road segmentation approach. The textures in road images are usually strongly anisotropic with a dominant orientation. Such oriented textures can be described by their orientation field that consists of orientation and strength of texture anisotropy. The authors describe a new approach for estimating the orientation field of oriented textures based on the covariance matrix of the grey value changes in an image. As an image feature, the strength of texture anisotropy is then used in the unsupervised road segmentation in the initial phase as well as in the supervised road segmentation in subsequent phases. The authors' algorithm has been tested with real road image sequences.