author = {Tom Botterill and Richard Green and Steven Mills},
title = {Voxel carving for collision avoidance for a vine pruning robot arm},
booktitle = {In Proceedings of Image and Vision Computing New Zealand},
year = {2014},
abstract={A vine pruning robot builds a 3D model of vines using a model-based feature matching and bundle adjustment pipeline. It is important to model the entire structure of each vine in order to plan collision-free paths for a robot arm, and to generate a connected vine structure. Each vine has a complex lump of old growth known as the head, which is challenging to model within the feature-based pipeline, so we use voxel carving to find the \textit{visual hull} of the head from many views. The visual hull is a volumetric model containing the head, so is ideal for collision avoidance. This paper describes how the voxel carving algorithm is integrated into the 3D vine modelling system. For our application it is convenient to model the vine head as a set of spheres, so we modify a standard voxel carving pipeline to to fit a set of spheres to images of the vine head. We evaluated the system on 300 stereo images of 15 vines: the voxel carve is computationally efficient, and the models found are over 90\% complete. }