Vision-Based Human Activity Recognition Methods Using Pose Estimation

Human Activity Recognition (HAR) has been a theme of great interest in research, especially thanks to the possible practical applications in the fields of video surveillance, human-machine interaction, gaming, autonomous driving, and health care. Despite this, the problem still remains a very complex challenge and whose definitive solution is still far away. Since the quality of the feature extracted impacts dramatically on the results of the overall system, the Human Pose Estimation (HPE) phase is crucial for the HAR. However, the problem remains of establishing how to use the extracted pose according to the network created for the HAR. Our work aims to analyze the ways in which the pose estimation, based on 2D poses extracted from monocular images, can affect a HAR system consisting of a recurrent network. For this purpose, various possible recurring structures are examined, at whose input the pose is used in different formats. The analysis carried out shows how the numerical simplification of the inputs facilitates learning compared to a “human” approach (which, on the contrary, could consider it easier to start from the graphic visualization of the skeleton).