Extracting Edge Voxels from 3D Volumetric Maps to Reduce Map Size and Accelerate Mapping Alignment

Abstract

For effective mobile robots we need a concise yet adequately descriptive mechanism for representing their surroundings. Traditionally 2D occupancy grids have proven effective for task such as SLAM, path planning and obstacle avoidance.Applying this to 3D maps requires consideration due to the large memory requirements of the resulting dense arrays. Approaches to address this, such as octrees and occupied voxel lists, take advantage of the relative sparsity of occupied voxels. We enhance the occupied voxel list representation by filtering out those voxels that are on planar sections of the environment to leave edge-like voxels. To do this we apply a structure tensor operation to the voxel map followed by a classification of the eigenvalues to remove voxels that are part of flat regions such as floors, walls and ceilings. This leaves the voxels tracing the edges of the environment producing a wire-frame like model. Fewer edge voxels require less memory and enable faster alignment.We compare the performance of scan-to-map matching of extracted edge voxels with that of the corresponding full 3D scans. We show that alignment accuracy is preserved when using edge voxels, while achieving a five times speedup and reduced memory requirements, compared to matching with all occupied voxels. It is posited that these edge voxel maps could also be useful for appearance based localisation.

A New Approach for 3D Edge Extraction by Fusing Point Clouds and Digital Images

“…Edge detection forms an important requirement for feature extraction in point clouds [3,4], which in turn can be divided in to two main categories [5]:…”