First Step Towards Embedded Vision System for Pruning Wood Estimation

This paper focuses on the development and evaluation of a portable vision-based acquisition device for vineyards, equipped with a GPU-accelerated processing unit. The device is designed to perform in-field image acquisitions with high-resolution and dense information. It includes three vision systems: the Intel® RealSense™ depth camera D435i, the Intel® RealSense™ tracking camera T265, and a Basler RGB DART camera. The device is powered by an Nvidia Jetson Nano processing board for both simultaneous data acquisition and real-time processing. The paper presents two specific tasks for which the acquisition device can be useful: wood volume estimation and early bud counting. Acquisition campaigns were conducted in a commercial vineyard in Italy, capturing images of vine shoots and buds using the prototype device. The wood volume estimation software is based on image processing techniques, achieving an RMSE of $2.1 ~cm^{3}$ and a mean deviation of $1.8 ~cm^{3}$. The buds detection task is obtained by fine-tuning the YOLOv8 model on a purposely acquired custom dataset, achieving a promising F1-Score of 0.79.