Normal distribution transform graph-based point cloud segmentation

We present a graph-based algorithm for segmenting point cloud scenes using criteria based on the combination of spatial, geometric, and appearance features. An octree data structure is employed to organize the point cloud data. The voxel space is used to create a Normal Distribution Transform Feature Representation (NDT-FR) to model the underlying sensor data and corresponding features in a probabilistic manner. The proposed segmentation algorithm uses the Hellinger distance calculated on local statistics stored in neighboring voxels to define the edge weights of the graph. Rather than choosing a specific feature for edge weight calculation, our approach has the ability to combine multiple features into a single edge weight without the need to find an appropriate normalization scheme. We verify our algorithm on multiple indoor scenes and perform a qualitative evaluation. We also show how our edge weighting scheme can increase the accuracy of object boundaries in the final segmentation.