Structure and Sequencing Preserving Representations for Skeleton-based Action Recognition Relying on Attention Mechanisms

Skeleton-based action recognition is a highly prominent research focus within the domain of computer vision, primarily owing to the availability and accessibility of skeletal data, and its stability in the presence of diverse transformations such as body scales, camera angles, and complex backgrounds. Modeling the change of temporal and spatial characteristics of skeletal data is essential in attaining this purpose. This paper proposes a range of representations of skeletal data and evaluates and contrasts them, first, to introduce distinct ways of simultaneously addressing temporal and spatial aspects, and second, to identify the most effective solution. Furthermore, to guarantee proficient training of the neural network, and based on the observation that the joints do not contribute equally to the execution of an action, we suggest strengthening the recognition task by incorporating attention mechanisms into the fundamental recognition structures, which is essential. To show the practicality of our developed representations, we conducted a series of experiments on the well-known UT-Kinect and Kinect Activity Recognition Dataset (KARD) datasets. Obtained results are comparable or better in terms of accuracy regarding previous work.