Multi-view Probabilistic Segmentation of Pome Fruit with a Low-Cost RGB-D Camera

Fruit harvesting is a topic of intereset in agricultural industries. In order to perform this task, robots should be able to recognise and segment fruit in their perceptual environment. Particularly, apple trees are often arranged as planar trellis structures in commercial orchards. The vine-like branches have leaves that can occlude fruit and produce noise in typical depth sensor data that also populates the scene with objects that are not of interest. In this paper, we present a method that uses a Dirichlet mixture of Gaussian processes and a Gibbs-Sampler for segmenting clusters of apples to support selective autonomous harvesting. Furthermore, the model provides probabilistic reconstruction of the entire apple which can be used for better grasping of the fruit.