High-precision 3D detection and reconstruction of grapes from laser range data for efficient phenotyping based on supervised learning

•Fully-automated 3D reconstruction of grape bunches for phenotyping.•Combining supervised learning and object recognition methods.•Parameter initialization and optimization based on known statistical values.•In-depth evaluation of all steps of the pipeline.

In this contribution, we present an automated approach to the phenotyping of grape bunches. To do so, our method analyses high-resolution sensor data taken from grape bunches and generates complete 3D reconstructions of the observed grape bunches. We extend a previous work from our group to earlier development stages with mostly visible stem structure, using an enhanced pre-classification of the sensor data into specific categories, i.e., berries and stems, yielding high precision and recall rates for the reconstruction of the berries of more than 98% and 94%, respectively. The same quality of results can be achieved by training a classification model on one grape bunch and applying it to the other grape bunches. Furthermore, we describe important observations concerning parameter initialization and optimization techniques resulting in a guideline for people working in the area.

Graphical abstractDownload high-res image (187KB)Download full-size image