This repository contains the code, data and plots of the project that I am implementing for openIoT Lab (FBK).
The OpenIoT research group has a ROS/ROS2-based mobile robot based on the JPL OpenRover open-source project. The rover carries a detachable sensing platform composed of an RGB camera and a 4 DoF robotic arm from the OpenManipulator series and is also ROS/ROS2-enabled. Using this platform, the final goal of the internship is to develop a ROS module that can use the camera to detect/count fruits. To ease the initial assessment of the developed system, an existent dataset comprising 1000+ still images of blueberries will be used.
In the src/python folder the user can find all the code necessary to generate the obtained results. Results can be obtained by running
python src/python/softDetection.py
The ML models that are contained in models have been created and tested in Edge Impulse. The detection performances has been calculated by running
python src/python/performances.py
a boolean variable named "bunchDetection" is used here to evaluate the performances of the desired task.
The use of a virtual environment for the installation is suggested as a common good programming choice. For example, the use of pipenv requires the following commands
pipenv shell
pipenv install -r software/python/requirements.txt
The detection accuracy calculated for the berry detection is of the 83.9% while the one regarding the bunch detection reaches the 50%, considering an IoU > 0.5 as a threshold for the true detection. Increasing the IoU threshold to 0.7 the precisions become 81% and 37.5% for berries and bunches respectively.
The data contained in the "dataset" folder have been taken from the Deepblueberry dataset for research purposes.
