Proceedings Article | 17 May 2013
KEYWORDS: Feature extraction, LIDAR, Clouds, 3D modeling, Robots, Sensors, Head, Data modeling, Data acquisition, Global Positioning System
This study explained a method to classify humans and trees by extraction their geometric and statistical features in data
obtained from 3D LADAR. In a wooded GPS-denied environment, it is difficult to identify the location of unmanned
ground vehicles and it is also difficult to properly recognize the environment in which these vehicles move. In this study,
using the point cloud data obtained via 3D LADAR, a method to extract the features of humans, trees, and other objects
within an environment was implemented and verified through the processes of segmentation, feature extraction, and
classification. First, for the segmentation, the radially bounded nearest neighbor method was applied. Second, for the
feature extraction, each segmented object was divided into three parts, and then their geometrical and statistical features
were extracted. A human was divided into three parts: the head, trunk and legs. A tree was also divided into three parts:
the top, middle, and bottom. The geometric features were the variance of the x-y data for the center of each part in an
object, using the distance between the two central points for each part, using K-mean clustering. The statistical features
were the variance of each of the parts. In this study, three, six and six features of data were extracted, respectively,
resulting in a total of 15 features. Finally, after training the extracted data via an artificial network, new data were
classified. This study showed the results of an experiment that applied an algorithm proposed with a vehicle equipped
with 3D LADAR in a thickly forested area, which is a GPS-denied environment. A total of 5,158 segments were
obtained and the classification rates for human and trees were 82.9% and 87.4%, respectively.
