Tracking unfragmented and slightly fragmented particles in a mini-triaxial sample using the neural networks

In this paper, a novel particle tracking method is proposed to investigate the kinematics of particles exhibiting zero or low-level fragmentation by combining PointConv and PointNetLK networks. Firstly, a series of image processing algorithms were employed on the 2D slices to facilitate reproduction of particle morphology and motion. Subsequently, all particles were represented as point sets, downsampled and grouped, yielding substantial datasets for neural network training and testing. Additionally, Gaussian noise was generated and introduced into the particle point sets to enhance the network robustness. Then, the PointConv network was implemented to efficiently track unfragmented and slightly fragmented particles under different strains. This success was attributed to the good preservation of morphological features of these particles and the excellent capacity of PointConv to capture morphological features. Next, the PointNetLK network was trained and integrated with the particle correspondence obtained by PointConv, aiming to determine the optimal transformation matrix for the corresponding particles. The predictive results are evaluated based on the visualization of the transformed point cloud, followed by a discussion of the influence of different parameters on the predictions. Finally, a comprehensive comparison with several existing particle tracking methods highlights the advantages of the proposed method.