Estimation of position and orientation with optic sensors in an industrial environment
Heino, Konsta (2022)
Heino, Konsta
2022
Automaatiotekniikan DI-ohjelma - Master's Programme in Automation Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. Only for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2022-09-27
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202209126989
https://urn.fi/URN:NBN:fi:tuni-202209126989
Tiivistelmä
Recent advancements with neural networks have made the six degrees of freedom pose es timation of objects with optic sensors more accurate and reliable. Accurate pose estimation is a critical factor in automating processes and creating safety applications. In this thesis, research is conducted and applied to address the problem: is it possible to design and build a real-time sys tem to localize, and estimate the pose of, an indoor crane hook in a manufacturing environment. To answer this, the needs of the application were analyzed, and a multi-part pipeline was de signed in accordance. Hardware and algorithms were researched and attained to gather RGBD data, this data is fed to a model and the model was created by training it with a dataset of the pose estimation target. Each section of the pipeline was designed and developed based on the data it takes in, the data it passes forward, and the pre-existing conditions of the application in this manufacturing use-case, with the eventual goal of successful pose estimation. After this phase the pose estimation pipeline was built, and the results were analyzed.
Individual sections of the pipeline were successes. However, the ultimate goal of pose esti mation could not be achieved during this research. The system provided good quality depth data with two different methods: stereo imaging and LIDAR. Further, applications were created for dataset creation, with both synthetic and real-life data, in accordance with current scientific re search on the subject. However, when problems arose with the trained pose-estimation model, there was no time to create another dataset, and the whole pipeline could not be tested without. In conclusion, the possibility of pose-estimation in the crane working environment could not be confirmed, but since the individual sections of pipeline performed well, it is promising in theory.
Individual sections of the pipeline were successes. However, the ultimate goal of pose esti mation could not be achieved during this research. The system provided good quality depth data with two different methods: stereo imaging and LIDAR. Further, applications were created for dataset creation, with both synthetic and real-life data, in accordance with current scientific re search on the subject. However, when problems arose with the trained pose-estimation model, there was no time to create another dataset, and the whole pipeline could not be tested without. In conclusion, the possibility of pose-estimation in the crane working environment could not be confirmed, but since the individual sections of pipeline performed well, it is promising in theory.