Abstract: Vision-based road detection is a popular area in research of driving security, however, detecting in complex road scenery is still a challenging topic. An approach is proposed to detect drivable road region from monocular images in urban environments. The algorithm is based on multi-scale sparse representation, with local texture in large scale, and context in medium scale. Experiments show that, distinguishing the similar texture of pavements from that of surrounding buildings and obstacles brings a well-performance in structured roads as well as the diverse road environments such as lack of lanes or clear boundaries but full of complex illuminations.