Feasibility Study of In-Field Phenotypic Trait Extraction for Robotic Soft-Fruit Operations

Abstract

There are many agricultural applications that would benefit from robotic monitoring of soft-fruit, examples include harvesting and yield forecasting. Autonomous mobile robotic platforms enable digitisation of horticultural processes in-field reducing labour demand and increasing efficiency through con- tinuous operation. It is critical for vision-based fruit detection methods to estimate traits such as size, mass and volume for quality assessment, maturity estimation and yield forecasting. Estimating these traits from a camera mounted on a mobile robot is a non-destructive/invasive approach to gathering qualitative fruit data in-field. We investigate the feasibility of using vision- based modalities for precise, cheap, and real time computation of phenotypic traits: mass and volume of strawberries from planar RGB slices and optionally point data. Our best method achieves a marginal error of 3.00cm3 for volume estimation. The planar RGB slices can be computed manually or by using common object detection methods such as Mask R-CNN.