Vision-based tracking of orientation and velocity for non-spherical particles

Abstract

A stereo camera method for tracking the position and orientation of rigid non-spherical particles is presented. The method works by comparing the images of the observed particle with that of a geometric model of the observed particle. The comparison of images is formulated as an optimization problem that determines the position and orientation of the particle at each time step. The optimization problem is solved using gradient descent by implementing the objective function in the differentiable programming library PyTorch. The method is validated using synthetic data, demonstrating robustness to both regular and irregular particle shapes. The influence of image noise, pixel density, and the resolution of the geometric model on the accuracy of the recovered position and orientation is addressed, and the results show sub-pixel accuracy to around 0.3 pixels at realistic experimental conditions. Finally, the method is applied to an experiment of non-spherical particles settling in a quiescent flow. Three regular and ten irregular particles are used, with particle Reynolds numbers ranging between 200 and 800. The results show that the instantaneous vertical velocity of the particles can vary by up to 50% due to changes in orientation. Different settling modes are identified, highlighting the importance of tracking all 6\(^\circ\) of translational and rotational freedom to capture the dynamics of settling non-spherical particles.