Diram Tabaa

High-fidelity 3D simulation is critical for training mobile robots, but its traditional reliance on mesh-based representations often struggle in complex environments, such as densely packed greenhouses featuring occlusions and repetitive structures. Recent neural rendering methods, like Gaussian Splatting (GS), achieve remarkable visual realism but lack flexibility to incorporate fiducial markers, which are essential for robotic localization and control. We propose a hybrid framework that combines the photorealism of GS with structured marker representations. Our core contribution is a novel algorithm for efficiently generating GS-based fiducial markers (e.g., AprilTags) within cluttered scenes. Experiments show that our approach outperforms traditional image-fitting techniques in both efficiency and pose-estimation accuracy. We further demonstrate the framework’s potential in a greenhouse simulation. This agricultural setting serves as a challenging testbed, as its combination of dense foliage, similar-looking elements, and occlusions pushes the limits of perception, thereby highlighting the framework’s value for real-world applications.

Advancements in 3D sensing technologies have led to an increased reliance on point cloud data for diverse applications ranging from autonomous navigation to environmental modeling. However, the sheer volume of data collected by these technologies poses significant challenges for real-time processing and analysis. This thesis introduces SampleLapNet, a novel neural network architecture designed to address the challenges of point cloud downsampling in a task-agnostic manner. By leveraging the Laplacian operator as a geometric measure of point importance, SampleLapNet learns to predict and preserve critical geometric features during the downsampling process, thereby ensuring minimal loss of relevant information. The architecture combines the robustness of transformer models with the efficiency of Laplacian-based importance scoring to facilitate efficient preprocessing that enhances subsequent point cloud analyses. We demonstrate the effectiveness of SampleLapNet through extensive experiments on benchmark datasets, showing significant improvements in downsampling efficiency without compromising the performance of downstream tasks such as semantic segmentation. This work not only proposes a method to reduce computational demands but also provides insights into the geometric processing of 3D data, suggesting pathways for future innovations in point cloud processing.