Analysis of data structures used for storing and processing 3D LADAR data

This paper compares the use of a point cloud data storage structure with a voxel based storage structure for 3D data collected with LADAR. The motivation for this work is to support the development of a classification system to model the environment of an autonomous vehicle operating in an unstructured natural setting. The classifier should be able to detect trees, bushes, and ground based on the data. Most LADAR sensors provide accurate 2-D range information in a plane, but we generate 3-D data from the sensor by articulating it using a pan-tilt mechanism. A 3-D grid of voxels is used with each voxel representing a cubic region in space with an associated value to indicate the region's occupancy. We discuss how this voxel representation compares with a point cloud representation in the process of data collection, data storage, and data processing and also look at the complexity of merging new data with old data in the same region. There are two main contributions. The first is comparing a possible voxel based data representation with one possible point cloud data representation. The second is determining reasonable parameters (voxel size, LADAR pitch angular velocity, and distance to the target) to facilitate classification of the voxel data. The scope of this paper is confined to analyzing data size, scanning time required for accurate classification, and processing efficiency. Experimental data size, density, and classification performance results are presented.