Design of a system for detecting and targeting bunches of grape and their petioles

This thesis treats an alternative method to perform an automatic harvesting of the grape bunches. The project tries to imitate the manual harvesting as much as it can, so that the bunch of grape is safely collected without being damaged. The thesis is divided into eight chapters, each of which treats a main topic. The paper begins with an introduction, in which the story of the harvesting of the grapes is shortly reported, along with the reasons why the current mechanical harvesting techniques can be improved in some respects. Then, the thesis treats the three main issues of this project, starting from the recognition of the bunch of grape, going to the recognition of the petiole of the just found bunch of grape, ending up to the targeting of that petiole in the space. For each one of these tasks, the procedures which have led to the final results are explained. The discarded solutions are reported too, along with the reasons why they have failed. The overall project is then assembled: the solutions of the issues previously treated are merged together and tested. The limitations imposed by the current design and the possible upgrades that could be implemented in order to boost the performances are also discussed. Since one of the main task, the grape bunches recognition, is based on both the neural networks and tensorflow, these two topics are treated in detail in two appendixes.