Point cloud data is received and a ground plane is segmented. A two-dimensional image of the segmented ground plane is generated based on intensity values of the segmented ground plane. Lane marking candidates are determined based on intensity within the generated two-dimensional image. Image data is received and the generated two-dimensional image is registered with the received image data. Lane marking candidates of the received image data are determined based on the lane marking candidates of the registered two-dimensional image. Image patches are selected from the two-dimensional image and from the received image data based on the determined lane markings. Feature maps including selected image patches from the registered two-dimensional image and received data are generated. The set of feature maps are sub-sampled, and a feature vector is generated based on the set of feature maps. Lane markings are determined from the generated feature vector.