Improving Action Recognition through Pose Estimation and Directed Graph Neural Networks

To improve the accurate performance and robustness of action recognition systems, this study attempts to merge the posture estimation and skeleton-based action recognition. While pose estimation provides a spatial representation of important places on the human body, skeleton-based action identification focuses on examining the time dynamics of human motions. By combining these two methods, we increase the performance of action recognition by utilizing the complementary data from both the spatial and temporal domains. Through empirical evaluation on benchmark dataset, this study shows the efficiency of alternative approaches for fusing posture estimation with action recognition. The outcomes show that this integrated method has the potential to attain cutting-edge performance in identifying human actions.