Fruit Localization using Artificial Vision and Shape Analysis

Abstract

This paper details a procedure for identifying edges that may belong to three-dimensional convex objects of an approximate size, such as apples, from a much larger set of edges. The identification is achieved by analyzing a number of intensity profiles that originate at each edge and determining whether they have a shape that is suitable with a 3D convex object of the correct size. We show that contrarily to the prevailing opinion, the intensity or luminance functions of three-dimensional convex objects are not necessarily convex, which led us to developing models for describing such profiles. The simplest suitable model includes four parameters that can be easily estimated by a standard constrained least squares optimization procedure. The proposed procedure is applied to images of apple trees recorded in an orchard. The procedure correctly removes 77-100% of the edges not belonging to apples, with minimal loss of edges belonging to such objects.