The PointNet architecture is a foundational deep learning model for 3D point clouds, solving classification and segmentation tasks. We hypothesize that the full potential of PointNet has not been reached and is greatly restrained by a single Max pooling layer. First, this thesis introduces new and more complex learnable aggregation functions. Secondly, a novel normalization technique, Context Normalization, is proposed for further feature extraction. Context Normalization is similar to Batch Normalization but independently normalizes each point cloud within a mini-batch and always uses dynamic statistics. The experiments show that replacing Max pooling with Principal Neighborhood Aggregation (PNA) increased classification accuracy from 73.3% to 78.7% on an SO(3) augmented version of the ModelNet40 dataset. Combining PNA with Context Normalization further increased accuracy to 84.6%.

<p>The LiDAR data of the maize used in this research were acquired from different stages, by different sensors, and from different flight heights, which results in different point densities. The ground reference data collected by LiCOR LAI-2200 represented the leaf area index of a two-row plot.</p>

Computational fluid dynamics (CFD) is the analytical process of predicting fluid flow, mass transfer, chemical reactions, and other related phenomena during the design or manufacturing process. Aggressive use of CFD provides drastic reductions in wind tunnel time and lowers the number of experimental rig tests. CFD saves hundreds of millions of dollars for industries, governments, and national laboratories, offering the potential to deliver superior understanding and insight into the critical physical phenomena limiting component performance. Thus, CFD opens new frontiers in many fields, especially vehicle design. One key strength of CFD is its ability to produce simulations useful in inverse design and optimization problems. However, a simulation in a conventional solver is considerably time-consuming to converge. To enable more efficient and scalable CFD simulations, we leverage the universal approximation property of machine learning using deep neural networks (DNNs) to estimate a surrogate solution to the CFD simulation. We present an implementation of this idea in two different models, one representing the eulerian model for compressible viscous flows and another representing the compressible Navier–Stokes equations. Lastly, we discuss the compressible Navier–Stokes network’s performance by implementing an inverse design problem to know if a gradient descent step of the model w.r.t the shape would grant the optimal solution. After training, predictions from these networks are faster than conventional solvers. The network predicts the flow fields hundreds of times faster than current conventional CFD solvers while maintaining good accuracy. Using the network’s predicted solutions to initialize a CFD solver sufficiently speeds up the simulation.

Online house building tools are robust instruments that have transformed the approach to home planning and design. The significance of 3D models on online house building and design platforms lies in their ability to elevate the user experience, enhance design precision, and foster collaboration. The creation of 3D house models is a demanding and detail-oriented undertaking that requires significant dedication and expertise. This thesis plans to investigate possibilities of implementing AI with a limited amount of data, to help streamline the manual process of working on 3D models for use in online house building tools. Multiple problems arouse related to issues with the acquired house mesh data, issues with the chosen AI models and hardware issues.  The problems during the different stages of work made any meaningful analyses hard but some evaluations could still be extracted. Time was a central aspect in that the time taken to train the AI models greatly correlated with the amount of data used for said training. It was also made apparent that there was a great need for more data if the AI was to be able to be trained on a dataset made from 3D house meshes. If this type of project was to be picked up in the future, a recommendation would be to not go at it alone as the time needed to perform the different steps should not be underestimated.

In this thesis we explore an infrared light-based approach to the problem of pose estimation during aircraft landing in urban environments where GPS is unreliable or unavailable. We introduce a novel fiducial constellation composed of sparse infrared lights that incorporates projective invariant properties in its design to allow for robust recognition and association from arbitrary camera perspectives. We propose a pose estimation pipeline capable of producing high accuracy pose measurements at real-time rates from monocular infrared camera views of the fiducial constellation, and present as part of that pipeline a data association method that is able to robustly identify and associate individual constellation points in the presence of clutter and occlusions. We demonstrate the accuracy and efficiency of this pose estimation approach on real-world data obtained from multiple flight tests, and show that we can obtain decimeter level accuracy from distances of over 100 m from the constellation. To achieve greater robustness to the potentially large number of outlier infrared detections that can arise in urban environments, we also explore learning-based approaches to the outlier rejection and data association problems. By formulating the problem of camera image data association as a 2D point cloud analysis, we can apply deep learning methods designed for 3D point cloud segmentation to achieve robust, high-accuracy associations at constant real-time speeds on infrared images with high outlier-to-inlier ratios. We again demonstrate the efficiency of our learning-based approach on both synthetic and real-world data, and compare the results and limitations of this method to our first-principles-based data association approach.

Introduction: Number of strata (NoS) is an informative descriptor of forest structure and is therefore useful in forest management. Collection of NoS as well as other forest properties is performed by fieldworkers and could benefit from automation. Objectives: This study investigates automated prediction of NoS from airborne laser scanned point clouds over Swedish forest plots.Methods: A previously suggested approach of using vertical gap probability is compared through experimentation against the geometric neural network PointNet++ configured for ordinal prediction. For both approaches, the mean accuracy is measured for three datasets: coniferous forest, deciduous forest, and a combination of all forests. Results: PointNet++ displayed a better point performance for two out of three datasets, attaining a top mean accuracy of 46.2%. However only the coniferous subset displayed a statistically significant superiority for PointNet++. Conclusion: This study demonstrates the potential of geometric neural networks for data mining of forest properties. The results show that impediments in the data may need to be addressed for further improvements.

Detecting defects in industrially manufactured products is crucial to ensure their safety and quality. This process can be both expensive and error-prone if done manually, making automated solutions desirable. There is extensive research on industrial anomaly detection in images, but recent studies have shown that adding 3D information can increase the performance. This thesis aims to extend the 2D anomaly detection framework, PaDiM, to incorporate 3D information. The proposed methods combine RGB with depth maps or point clouds and the effects of using PointNet++ and vision transformers to extract features are investigated. The methods are evaluated on the MVTec 3D-AD public dataset using the metrics image AUROC, pixel AUROC and AUPRO, and on a small dataset collected with a Time-of-Flight sensor. This thesis concludes that the addition of 3D information improves the performance of PaDiM and vision transformers achieve the best results, scoring an average image AUROC of 86.2±0.2 on MVTec 3D-AD.