Ing. Jan Glaser

Ing. Jan Glaser

Ing. Jakub Novák

Department of Applied Mathematics

This thesis presents a playing card recognition system that uses two approaches. The first approach is based on computer vision techniques to classify playing cards in a controlled envi- ronment. The second approach involves creating a comprehensive dataset with augmentation and training a deep-learning model. The first approach uses image processing techniques to detect the rank, suit, and color of a single playing card in the image with multiple methods. The best method achieves an accuracy of 100% for rank and suit identification and 78.16% for color identification on a dataset of 728 playing cards. The second approach involves creating a comprehensive dataset of playing cards with augmentation techniques. Initially, the dataset contains 30,000 images of playing cards labeled with their rank and suit, and these images are subjected to slight augmentations. The YOLOv8 model is pre-trained on this initial dataset in the first phase of the training process. In the second phase, the model is fine-tuned on a larger dataset of 80,000 images, which includes multiple cards in each image and more closely resembles real-world conditions, such as different points of view and varying lighting conditions. The model is also validated during the fine-tuning process to ensure optimal performance. Once the model is trained, it is tested on a set of 30,000 images that maintain the diversity and complexity found in real-world environments. The YOLOv8 model achieves an accuracy of 99.98% on this test set, demonstrating its effectiveness in recognizing playing cards in a wide range of environments.

Thesis on DSpace

Ing. Jan Glaser

Ing. Ondřej Smrž

Department of Applied Mathematics

This thesis delves into developing a pipeline for registering partially overlapping large-scale point clouds and measuring the influence of semantic label information in the registration process. The result of this work is a method that uses the additional information obtained from convolutional neural networks to partition the point clouds and align local structures using principal component analysis. This method was tested against traditional methods on publicly available datasets. Experiments have shown that aligning smaller subsets proved much more stable and accurate than traditional approaches.

Thesis on DSpace

Ing. Jan Glaser

prof. Ing. RNDr. Martin Holeňa, CSc.

Department of Applied Mathematics

The aim of this thesis is to design a neural network for reconstructing missing parts in 3D point clouds and to evaluate how this reconstruction affects the accuracy of point cloud registration. The proposed network predicts only the missing points and concatenates them with the original input, leaving the observed geometry unchanged. Experiments have shown that completion significantly improves registration, especially with moderate amounts of missing data, where the model has enough remaining structure to recover the missing geometry.

Thesis on DSpace

Ing. Jan Glaser

prof. RNDr. Pavel Surynek, Ph.D.

Department of Applied Mathematics

The aim of this thesis is to automate the powder pre-spraying process in industrial powder coating for objects with complex three-dimensional geometries. The proposed approach operates on a CAD model of the coated object and focuses on identifying surfaces that are insufficiently covered during standard automated powder application. Several geometric analysis methods are evaluated, including oriented bounding box raycasting and surface curvature analysis, to detect areas requiring pre-spraying. Based on the detected regions, a method for generating robotic spray waypoints and trajectories is implemented using voxel-based sampling and path ordering. The resulting trajectories are demonstrated in simulation, showing the applicability of the proposed CAD-based framework for automated powder pre-spraying of complex geometries.

Thesis on DSpace