An Experimental Study on Pose Estimation Models and Data Preparation Strategies for Human Posture Assessment Pipeline

Abstract

Human posture assessments play a crucial role in improving quality of life and performance, and even in preventing injuries; however, real-world applications still face many challenges. Multiple people per image, varying camera angles, and varied person orientations are recognized as existing obstacles in academia. This paper gives a structured exploration of four development trials completed while developing a comprehensive posture- analysis pipeline. Firstly, a representation of a comparative analysis of five popular human datasets: 3.6M, MPII, COCO, CMU Panoptic, and Human Eva. Secondly, four different human pose estimation models—Open Pose, MediaPipe, MMPose RTMO-I, and YOLO- NAS-POSE—were evaluated against predefined criteria after testing on the same sample of images. Thirdly, data preprocessing techniques—class imbalance handling, feature engineering, and splitting order—were applied on a custom-built and labeled dataset for the view classification module. Finally, a Unity-generated avatar dataset with normal, conditional normal, and abnormal human posture representations. The results of those four trials guided the design choices for the final pipeline. RTMO-I was selected as the preferred multi-person HPE model, where accuracy was chosen over latency. None of the five datasets used as image datasets; instead, a custom dataset was built from the COCO dataset annotation files and labeled with custom rules. Class imbalance was managed internally by the classification model; feature engineering was left out for simplicity, as that distance-related features does not improve the performance of our model, and formal splitting of the dataset before preprocessing steps was preserved. Finally, low-contrast, standing-still side-view avatars were generated to test el- bow and knee posture for verification. These findings emphasize practical considerations in designing real-world posture-analysis systems. Additionally, they indicate the benefit of iterative methodological trials in refining both model selection and dataset design.