Repository for "Populating Digital Orchards: Evaluation of a 3D Positioning Algorithm for Semi-Automated Data Collection and spawning", Simini et. al., IEEE MetroAgriFor 2024, Padova (IT).

Abstract
Precision Orchard Management tasks require high-grade precision representation of plants information to unleash the potential and the efficiency of involved operations. In order to improve this, especially when autonomous vehicles are involved, a 3D orchard reconstruction is particularly useful. This work aims to exploit RGBD images to estimate the local 3D position of fruits on a plant, with respect to the trunk position. Besides being a useful information to map plant production distribution, it can also be used to localize real objects in virtual environments, such as digital twins. The process can serve as a foundation of numerous monitoring activities, thinning procedures and autonomous harvesting tasks.

---

Files Nomenclature

Files are named using the following structure:

Each component in the filename has a specific meaning:

• <#img>: Represents the file identifier. This is assigned to color, depth annot according to the file content.
• <trial>: Indicates the trial number. It represents the specific experiment or test case in which the image was used.
• <repetition>: Refers to the repetition number. This indicates how many times a specific trial has been repeated to account for variations and ensure accuracy.
• <background>: Specifies the background condition or setting used during the test. It helps distinguish tests conducted under different environments or configurations (values: 0 leafy, 1 black background).

Example:
For a file named color_1_1_0.png, it can be interpreted as:

• color: RGB image
• 1: Trial number 1
• 1: First repetition
• 0: Leafy background

For a file named depth_1_1_0.png, it can be interpreted as:

• depth: Grayscale depth map
• 1: Trial number 1
• 1: First repetition
• 0: Leafy background

For a file named annot_1_1_0.json, it can be interpreted as:

• annot: file including manual annotations made on the color image
• 1: Trial number 1
• 1: First repetition
• 0: Leafy background

A more detailed description of files nomenclature can be found here

Data Processing and Analysis

All the code used for data Processing and Analysis are available in the notebooks directory divided for trial type and setup.

More information can be found in the conference proceedings paper (IN PRESS - Will be linked here once released) and in the presentation

Results

Laboratory test

Here below some graphical output obtained from the developed algorithm for the different trial

artificial tree (2D setup and leafy background)

picture	reference twin	estimated twin

artificial tree (2D setup and black background)

picture	reference twin	estimated twin

artificial tree (3D setup and leafy background)

picture	reference twin	estimated twin

artificial tree (3D setup and black background)

picture	reference twin	estimated twin

Field trial

Left- Mean Bias Error (MBE) per Axis; Right - Cumulative Error per Position

planar cordon (2D setup and leafy background)

spindle-bush (3D setup and leafy background)

Positioning performance evaluation

Laboratory trial

A 2D architecture demonstrated in all lab and field conditions to reduce the positioning error.

Errors obtained using the artificial tree.

Correlation heatmap showing no effect of object size on positioning.

---

Get in contact with AUTHORS