Detecting objects in cargo handling operations from 3D point cloud data
Siltanen, Janne (2022)
Siltanen, Janne
2022
Automaatiotekniikan DI-ohjelma - Master's Programme in Automation 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
2022-05-10
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202204263640
https://urn.fi/URN:NBN:fi:tuni-202204263640
Tiivistelmä
The objective of this master thesis is to research existing 3D point cloud deep learning methods and their ability to detect keypoints in automatic cargo handling operations. Point clouds contain spatial information, which is one reason why they have gained popularity in automated operations in many industries. Deep learning can be utilized to detect points of interest from all kinds of data and research on detection from point clouds has increased recently.
This master thesis researches the usage of convolutional and transformer-based deep neural networks for detecting key points from point cloud data. Six different point cloud deep learning methods will be covered and four different methods will be selected and utilized for testing with the data from automatic cargo handling operations. The methods will be tested for inference time and accuracy and compared to the method currently used in the production system. Inference time will be tested on edge environments and training server environment.
The thesis first presents the background of the problem in the introduction along with the research questions. Theoretical background of the related topics will be presented. Deep learning with point clouds has its own characteristics and problems when compared to 2D images so those will be covered in the theoretical part as well. This thesis will also cover how the training is executed, what kind of tests are used and how the used scripts operate.
The results show that in terms of accuracy convolutional-based method performed the best out of the researched methods and it outperformed the currently used method. Good results were also obtained with transformer-based models. In terms of inference time, not all methods can be utilized in an edge machine learning environment unless GPU computation is available.
This master thesis researches the usage of convolutional and transformer-based deep neural networks for detecting key points from point cloud data. Six different point cloud deep learning methods will be covered and four different methods will be selected and utilized for testing with the data from automatic cargo handling operations. The methods will be tested for inference time and accuracy and compared to the method currently used in the production system. Inference time will be tested on edge environments and training server environment.
The thesis first presents the background of the problem in the introduction along with the research questions. Theoretical background of the related topics will be presented. Deep learning with point clouds has its own characteristics and problems when compared to 2D images so those will be covered in the theoretical part as well. This thesis will also cover how the training is executed, what kind of tests are used and how the used scripts operate.
The results show that in terms of accuracy convolutional-based method performed the best out of the researched methods and it outperformed the currently used method. Good results were also obtained with transformer-based models. In terms of inference time, not all methods can be utilized in an edge machine learning environment unless GPU computation is available.