Bachelorarbeit BCLR-2020-54

Schnell, Miriam: Plant detection and classification in agricultural fields.
Universität Stuttgart, Fakultät Informatik, Elektrotechnik und Informationstechnik, Bachelorarbeit Nr. 54 (2020).
64 Seiten, englisch.

Field game herbs like the field larkspur provide a food source for animals e.g. bees, especially during their blooming period. This plant species is weak in competition for nutrients, thus it does not decrease crop yield. Nevertheless, the field larkspur is destructed by the field-wide application of plant protection products. The goal of this thesis is to explore different object detection, classification and segmentation approaches for their eligibility as crop-weed-discriminators for deployment in an autonomously acting weeding robot to be able to omit harmless plant species like the field larkspur during the weeding process. Prior to the development, a fundamental literature research is conducted and the state-of-the- art methods are summarised and compared. Subsequently the task is precisely defined and the requirements are derived, since they form the foundation for following assessments. After determining the system architecture of relevant components and selecting the utilised software, the first experiments are conducted. This includes fine-tuning of a pre-trained Mask R-CNN model on the Carrot-Weed data set and on three subsets of the Sugar Beets 2016 (SB16) data set to fulfil the vegetation detection, classification and segmentation task. Following the validation and verification of these experiments, the same pre-trained Mask R-CNN model is fine- tuned on the subset of SB16 data set comprising multi-class annotations with the goal to detect, classify and segment objects into the three classes crop, soil and weed. The weed class comprises all weed species instead of splitting them into different classes to enable the model to identify previously unknown weed types with less effort. Finally YOLOv4 is fine-tuned on the same multi-class annotated data set aiming to fulfil the crop-weed detection and classification task in real-time. Finally all methods of resolution are evaluated. With the findings of the present thesis it is possible to detect, classify and segment vegetation and background on agricultural fields and more importantly to distinguish the vegetation between crop and weed. This can be enhanced in future research by adding harmless non-crop plant species to the crop class, thus enhance the diversity on the agricultural field, which will also contribute to preservation of natural habitats for animals.

Volltext und
andere Links
Abteilung(en)Universität Stuttgart, Institut für Architektur von Anwendungssystemen
BetreuerAiello, Prof. Marco: Bregler, Kevin; Jordan, Florian
Eingabedatum18. Januar 2021
   Publ. Institut   Publ. Informatik