Automatic lane detection is an essential component for autonomous navigation systems. It is a challenging task in unstructured environments where lanes vary significantly in appearance and are not indicated by painted markers. This paper proposes a new method to detect pedestrian lanes that have no painted markers in indoor and outdoor scenes, under different illumination conditions. Our method detects the walking lane using appearance and shape information. To cope with variations in lane surfaces, an appearance model of the lane region is learned on-the-fly. A sample region for learning the appearance model is automatically selected in the input image using the vanishing point. This paper also proposes an improved method for vanishing point estimation, which employs local dominant orientations of edge pixels. The proposed method is evaluated on a new data set of 1600 images collected from various indoor and outdoor scenes that contain unmarked pedestrian lanes with different types and surface patterns. Experimental results and comparisons with other existing methods on the new data set have demonstrated the efficiency and robustness of the proposed method.