Computer vision frameworks for physics-based simulation of liquids and solids

Abstract

Simulating and visualizing fluid and solid materials in agricultural domains is an important and challenging problem in scientific computing and computer vision. Modern seed breeding programs require the ability to analyze seeds efficiently to be useful. Even simple measures such as volume and density can be challenging to compute efficiently with modest equipment. The dynamics of liquid and soil materials involve significant deformation during storm flows and require sophisticated numerical algorithms to achieve sufficient accuracy and visual realism. This dissertation focuses on extending volume carving techniques to measure seed volume and to create a new Material Point Method (MPM) models and finite volume models to simulate solids and fluids for dam safety analysis and visualization. This dissertation makes the following major contributions: The first is to create a novel framework for the design and analysis of computer experiments. The framework is applied to perform efficient dam breach and internal erosion analysis on a large number of structures. Given historical dam breach or design data input, the modeling framework can also be used to conduct sensitivity analysis to determine which parameters make the most impact on the resulting dam erosion. The second contribution is to develop new models for numerical simulation of dam erosion by combining fluid flow models developed using Computational Fluid Dynamics (CFD) with new dam erosion models using the Finite Element Method (FEM). A new model that combines fluid flow and erosion simulation into a single model is also developed using the Material Point Method (MPM). The third contribution is to build a comprehensive image capture and processing framework for seed property analysis. Rather than having a human manually measure seed properties such as length, width, thickness, and volume, the framework can automatically analyze a set of images from multiple angles and calculate the physical measurements for single seed samples. Finally, image analysis is extended using deep learning to increase the accuracy of rice image classification. The proposed frameworks are suitable for larger scale and more dynamic data in both dam safety and agricultural domains. They are also useful for computer animation in developing physics-based special effects for the animation of dam erosion. Previous work on MPM has resulted in models used in animation for Disney Studios, and the new models proposed could be used for accurate animation of fluid flows and dam erosion. Finally, the combination of image analysis algorithms and deep learning has many applications in the biomedical domain as well as the agricultural domain.