Nutzung der HoloLens 2 für die Pflanzenerkennung und Augmentierung im Kontext des EDEN ISS Weltraumgewächshauses

Kurzfassung

The aim of the following master thesis was to examine if the HoloLens 2 is able to perform plant object detection, which could be used to support the operating procedures of future astronauts inside space greenhouses with the goal to reduce crew time and workload. Review of related work in the field of machine learning and agricultural object detection indicated that the object detection framework YOLOv5 was suitable for such a usecase. Therefore a dataset containing images of wild arugula plants from the EDEN ISS analogue space greenhouse was compiled in order to train the YOLOv5 object detection model. During training different data augmentation methods where used to gradually improve the performance of the model. Subsequently a Unity application for the HoloLens 2 was developed to deploy the object detection model onto the HoloLens 2 and visualize the detection results with interactive 3D labels. The results of a performance test showed that the HoloLens 2 was not capable of performing a object detection on images with a resolution of 1920x1080 for a realtime application. Only when the resolution was reduced to 320x192 the application running on the HoloLens 2 could be described as usable. Furthermore a perspective test was conducted where the detection ability of the arugula model was compared to a pretrained YOLOv5 model in a real world environment using the HoloLens 2. The test consisted of recognizing a wild arugula plant with the arugula model from different heights and distances and doing the same with a similar sized object as the arugula for the pre-trained model. The results of this test indicated that the dataset which the arugula model was trained on was insufficient because the model had a worse detection range compared to the pre-trained model. The results of the performance test are limited by the fact that no real user evaluated the application for determining to what extend it is usable. Also the results of the perspective test are limited because it hadn’t been conducted inside the analogue space greenhouse which the training data was recorded in. Overall the application works as intended by detecting plants and annotating them accordingly. Because of the limitations a user study inside a greenhouse is necessary to determine how the model performs on the HoloLens 2 and how usable the application is in its current state. Finally the application needs to be tested and optimized further regarding the performance and detection range in order to provide a benefit for users in a real world environment.