Original scientific paper
https://doi.org/10.30765/er.3423
Multi-person 2D human pose estimation a benchmark for real-time applications
Tomislav Prusina
; School of Applied Mathematics and Computer Science, University of Osijek, Osijek, Croatia
*
Juraj Benić
; School of Applied Mathematics and Computer Science, University of Osijek, Osijek, Croatia
Domagoj Ševerdija
; School of Applied Mathematics and Computer Science, University of Osijek, Osijek, Croatia
Domagoj Matijević
; School of Applied Mathematics and Computer Science, University of Osijek, Osijek, Croatia
* Corresponding author.
Abstract
Human pose estimation (HPE) is a critical component of computer vision, enabling real-time applications across various fields. This study benchmarks six state-of-the-art frameworks: OpenPose, YOLO, RTMO, RTMPose, Sapiens, and MoveNet in the context of 2D multi-person pose estimation in videos for real-time applications. Using the Panoptic dataset and standardized metrics, including Object Keypoint Similarity (OKS), Average Precision (AP), and Average Recall (AR), we evaluate their accuracy, speed, and hardware efficiency under realistic conditions such as occlusion and crowded scenes. Our analysis highlights the strengths and limitations of each framework, providing valuable insights to help practitioners select and deploy reliable HPE solutions in real-world applications.
Keywords
benchmark; human pose estimation; MoveNet; OpenPose; sapiens; RTMO; RTMPose; YOLO
Hrčak ID:
350091
URI
Publication date:
4.8.2026.
Visits: 0 *