Deep learning based object detection with depth camera

The increase in the world population creates a need for new technologies in order to meet this increase in agricultural production. In this context, it is an obvious reality all over the world that there are problems in the harvesting part of agricultural production in terms of finding the necessary labor force and efficient harvesting processes. In this regard, the development of autonomous harvesting robots and their active use in the fields are presented as a solution to this problem. There are studies to develop autonomous harvesting robots for many products around the world. In terms of these studies, especially the developments in computer vision studies, the successful results obtained in deep learning algorithms used in object detection and classification have made it possible to actively use autonomous harvesting robots in agricultural fields. In addition to these developments in software, with the development of the cards on which these algorithms will work in hardware, the processing speed of autonomous harvesting robots increases and harvest times are reduced. With the development of autonomous harvesting robots, solutions are being developed to the problems of labor power in the agricultural field, human-induced errors and occupational accidents. In this thesis, a deep learning-based object detection study was carried out to be used in autonomous harvesting of kumquat fruit. In this context, the most suitable model was determined by training with four different architectures of the YOLO algorithm in order to detect real-time objects that autonomous harvesting robots need in the vision system. After the object detection, the collection point needed for the harvest of kumquat fruit was determined. With the collection point detection, it is aimed to prevent problems such as leaves and branches that will cause problems in the autonomous harvesting process. As a result of the test processes, the most successful results were obtained in the YOLOv7 model with 93% overall accuracy and 35 FPS (Frame Per Second).