Trajectory Augmentation Method Based on Dynamic Movement Primitives

In this paper, we introduce a new augmentation method that takes into account the inherent properties of trajectory data and regenerates valid trajectories while preserving all the distinctive features of the main path. Our method uses the Dynamic Movement Primitives (DMP) formulation, which is traditionally/widely used in path generation in robotics, to manipulate the data in a kinematically accurate way and employs the algorithm as an augmentation method for the first time. To this end, we applied the presented method to our Iranian sign language dataset by augmenting each group in our dataset with the proper form of our proposed DMP data augmentation method. After training our augmented dataset with two simple and widely used deep classification models, we improved our mean accuracy from 65.3% to 77.61% in one-shot learning and from 74.05% to 81.86% with two data from each class. On its own and combined with other methods, our method outperforms various other commonly used methods in the literature and can improve the classification and prediction results in various types of trajectory datasets.