| 摘要 |
Automatic scene parsing involving obtaining an image about an object of interest; obtaining a three dimensional (3D) point cloud about the object of interest 201; aligning the 3D point cloud with the image 202; segmenting the image into a plurality of superpixels preserving a graph structure and spatial neighbourhood of pixel data of the image 204; associating the superpixels in the image with a subset of 3D points representing a planar patch of the object 206; extracting a plurality of 3D features for each patch 208; and assigning a vector representing a 3D feature with a semantic label on the basis of the extracted 3D feature of the patch 210. Preferably, the point cloud is derived using LIDAR. The feature extraction may involve parameters such as the height or distance to the patch, its surface normal or the 3D point density of the patch. The semantic labeling step may use a training classifier algorithm such as a boosted decision tree. The exemplary application involves a panoramic camera and laser scanner attached to a vehicle for street mapping. |
