Pose Estimation of Industrial Pallets using Machine Learning : Monocular Color Camera Methodology and Execution
Koskelainen, Olli (2024)
2024
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-01-25
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2023122911237
https://urn.fi/URN:NBN:fi:tuni-2023122911237
Tiivistelmä
Object pose estimation is an area of research in the field of computer vision which is concerned with determining the position and orientation of objects from sensor data. The ability to accurately predict object pose is an important requirement in many popular research topics such as autonomous driving, robotics and augmented reality. Pose estimation has further applications in industries such as entertainment, medical care, agriculture, as well as security and surveillance. While pose estimation can be performed using different sensors, in this work we will focus on monocular RGB camera based methods.
The first major contribution of this thesis is to describe a methodology for creating a pose estimation model. We discuss data synthesis, data collection, annotation, techniques for estimating pose and metrics for measuring performance. We also propose a tool for annotating pose on large datasets. Special emphasis is placed on generating photorealistic data by leveraging modern rendering technology. This methodology is intended as a framework for implementing and evaluating a pose estimation model for any arbitrary object.
As our second contribution, we follow our methodology to create a pose estimator for logistics pallets. We design a pose estimation model based on previous work and test the performance impact of it's components through a process of elimination. We also compare performance of models trained on different dataset variations to find desireable qualities in synthetic datasets.
We collect and annotate real data which we use to evaluate our models. Our results serve as a proof of concept that a monocular RGB pose estimation solution for pallet detection is certainly possible, and even when the model is trained exclusively on synthetic data, it is able to generalize to the real world domain.
The first major contribution of this thesis is to describe a methodology for creating a pose estimation model. We discuss data synthesis, data collection, annotation, techniques for estimating pose and metrics for measuring performance. We also propose a tool for annotating pose on large datasets. Special emphasis is placed on generating photorealistic data by leveraging modern rendering technology. This methodology is intended as a framework for implementing and evaluating a pose estimation model for any arbitrary object.
As our second contribution, we follow our methodology to create a pose estimator for logistics pallets. We design a pose estimation model based on previous work and test the performance impact of it's components through a process of elimination. We also compare performance of models trained on different dataset variations to find desireable qualities in synthetic datasets.
We collect and annotate real data which we use to evaluate our models. Our results serve as a proof of concept that a monocular RGB pose estimation solution for pallet detection is certainly possible, and even when the model is trained exclusively on synthetic data, it is able to generalize to the real world domain.