Skip to Main Content
Recently, there are many autonomous navigation applications done in outdoor environment. However, safe navigation is still a daunting challenge in terrain containing vegetation. Thus, a study on vegetation detection for outdoor automobile navigation is investigated in this work. At the early state of our research, we focused on the segmentation of LADAR data into two classes by using local three-dimensional point cloud statistics. The classes are: scatter to represent vegetation such as tall grasses, bushes and tree canopy, surface to capture solid objects like ground surface, rocks or tree trunks. However, the only use of 3D features would never result a real robust vegetation detection system because of lacking color information. We, hence, propose a 2D-3D combination approach which can utilize the complement of three-dimensional point distribution and color descriptor. Firstly, 3D point cloud is segmented into regions of homogeneous distance. The local point distribution is then analyzed for each region to extract scatter features. Secondly, a coarse 2D-3D calibration needs to be implemented in order to map the regions to the corresponding color image. Then, color descriptors are studied and applied to each region and considered as color features. Those all scatter and color features will be trained by Support Vector Machine to generate vegetation classifier. Finally, we will show the out-performance of this approach in comparison with more conventional approaches.