Estimating 6D pose of an object using RGB data
Lampinen, Pauli (2022)
2022
Konetekniikan DI-ohjelma - Master's Programme in Mechanical Engineering
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2022-04-07
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202203302870
https://urn.fi/URN:NBN:fi:tuni-202203302870
Tiivistelmä
Estimating object’s 6D pose is an important aspect of automating even complicated processes, especially in a field of robotics where information about object’s 6D pose can be used for manipulating objects with a robot. In this thesis, various methods for estimating object’s 6D pose utilizing traditional approaches and machine learning are researched.
Possibility of estimating 6D pose of an object using RGB data is researched by utilizing a machine learning scheme. Data for training the machine learning model is simulation-based, heavily augmented for reaching best possible performance. Performance of machine learning architecture is tested on synthetic data as well as in a real-world industrial hall setting with real objects. In industrial hall setting, performance is evaluated by testing if the machine learning model can provide estimations with sufficient accuracy for successful robot grasping of an object. Even though machine learning model’s performance on synthetic data is reasonable, the machine learning model is not accurate enough for a reliable robot grasping operation in a real-world setting.
Possibility of estimating 6D pose of an object using RGB data is researched by utilizing a machine learning scheme. Data for training the machine learning model is simulation-based, heavily augmented for reaching best possible performance. Performance of machine learning architecture is tested on synthetic data as well as in a real-world industrial hall setting with real objects. In industrial hall setting, performance is evaluated by testing if the machine learning model can provide estimations with sufficient accuracy for successful robot grasping of an object. Even though machine learning model’s performance on synthetic data is reasonable, the machine learning model is not accurate enough for a reliable robot grasping operation in a real-world setting.