AbstractStrawberries are planted widely around the world and are an important cash crop, but strawberry farming has quite a high labour cost. Due to ageing populations and urbanization in many of the developed countries, as well as some extreme situations, such as the recent pandemic, there can be agricultural labour shortages. These factors motivate a strong need for a strawberry harvesting robot.
The basic process of automated strawberry harvesting could include tasks such as identifying mature strawberries ready to pick, controlling the robotic arm to approach the target, and removing the fruit from the plant. In contrast to hardware design, my research interests are more about the control system, especially about improving the efficiency of the robots. So far, the major bottlenecks in current strawberry harvesting robots are single image detection rates and the harvesting cycle times, especially in complex cases when strawberries are overlapping with each other or have obstacles like leaves nearby.
This thesis focuses on three key challenges: 1) identifying the strawberry picking point identification, 2) comparing detection methods, and 3) target sequence sorting when trajectory planning for strawberry harvesting. The aim is to reduce the effector of the bottlenecks identified above by improving the detection method and improve the targets sequencing method.
In the first part of the thesis, an updated picking point detection method includes the application of some machine learning methods are presented, which is based on an original method proposed during my MSc project. The results using different object detection method are compared and the detection accuracy working on different datasets are analysed.
In the next part, human preferences about the detection behaviours of robots are compared. Understanding human user preferences of their collaborators will not only help to improve the cooperation, but also provides some guidance when designing the target detecting system and the graphic user interface if necessary. Two groups of experiments are presented. In the first experiment, robot behaviours with different informative levels are compared; and in the second experiment, robot behaviours provided different amount of type I and type II errors are compared. Results showed that users trust both of the detection methods and are satisfied with their working speed in general, but do have preferences when working in different scenarios.
As the last part of the thesis, a target sequence sorting method learned from human users are introduced. In strawberry harvesting, cycle times can be improved by optimizing the harvesting sequence. A video game was created in order to collect data during the Covid-19 pandemic. The game simulates the arrangement of strawberries grown on a farm, and the game requires players to make decisions about harvesting. The data collected from players indicates that there is a pattern to human behaviour when planning the sequence in which strawberries are harvested. This target sequence sorting method is compared with some other harvesting sequence sorting methods usually applied to fruit harvesting robots. Moreover, this method is tested on an UR3 arm with plastic strawberry setups.
|Date of Award
|1 Jul 2023
|Simon Parsons (Supervisor) & Elizabeth Sklar (Supervisor)