Autonomous Pollination Application for Strawberries
The flowers of strawberries are insect-pollinated flowers that require insects to pollinate in the nature environment. However, insects will not be found in indoor hydroponic farms. Therefore, manual pollination is required since insects cannot be found in the indoor hydroponic lab.
In order to reduce the manpower and enhance the automation of urban indoor hydroponic farming, an autonomous pollination system will be developed for Vegetable Marketing Organization (VMO).
This application applies YOLOv4 for detecting the strawberry flowers and for classifying the color of anthers. Also, stepper motors, servo motor are using for controlling the movement of pollination gantry.
System Architecture
Pollination System
The system divide the rack into different grids, and the pollination gantry (camera and sweep) will go through every grid and detect for the strawberry flowers.
If the color of the anther is yellow, the pollination gantry will move
from the center of the grid to the space above the flower, and then it
will pollinate the flower. After that, the pollination gantry will move back to the centre of the grid and repeat the previous steps until all flowers are pollinated.
Flower Growing Tracker
The flower growing tracker applies coordinate matching on the flower coordinates obtained from YOLOv4 at every grid, and compares the previous and current coordinates of the strawberry flowers. This method is suitable for the pollination system which has been divided into different grids. It assumes that the detector detected three flowers in both the previous and the current state. Then, the coordinate matching calculates the Euclidean distance between every previous location and all new locations. After calculating the Euclidean distance, a combination of the distances that are the shortest from the set of old locations to the set of the new locations will be obtained, which is the new location of the flower. Therefore, the growing of the flowers can be tracked by coordinate matching.
Strawberry Flowers Detectiones
The YOLOv4-Tiny model used images collected from the VMO to train the model. Those images were divided into two groups, which are the training group and the validating group. Three classes have been labelled, which include “flower”, “yellow_anther”, and “brown_anther”.
Cross-validation has been applied for further accuracy testing of the Strawberry Flowers Detection Model. For the cross-validation, 461 images from the dataset have been equally divided into three groups (Group A, B and C). The average [email protected] IoU threshold result from the cross validation test is more than 80%, which is able to rep- resent the accuracy of the model.
Collaborations
Our Team
Sunny Sin - Mechanical Engineer
Leon Lam - Mechanical Engineer
Supervisor
Dr. Hayden K.H. So - Department of Electrical and Electronic Engineering, The University of Hong Kong
Dr. P. Lu - Department of Mechanical Engineering, The University of Hong Kong