International Journal of Information Technology and Applications, Vol. 1, No. 3, pp. 114-127, September 2024.
Abstract: The purpose of this study is to construct a dataset of human pose. In computer vision (CV), human pose estimation (HPE) is an important direction. After getting pose information, we can better analyze human motion and action. High-quality datasets are necessary during the work to train the HPE models. Nowadays, many kinds of datasets have been developed to meet these needs, and based on them, many methods have achieved excellent performance. However, pose estimation is still tricky for some unique and complex actions or motions, such as dance and sports actions. Aiming to solve this problem, we propose constructing a multiview-based pose dataset to analyze human action. In this work, we use filtering as an optimization approach to improve the accuracy of results. By filtering data, those wrongly recognized vital points can be corrected to some extent, and the missing key points or poses can be estimated based on the continuity of motion. The filter mothed can also improve the smoothness of the curve, so after filtering, the pose sequences can be more fluent and then conform to the kinetics. Sometimes, some errors in the 2D pose may result in images from some views caused by occlusions, environmental changes, and other reasons. In this situation, choosing 2D pose results from appropriate views can further improve the accuracy of the final results. Besides, for the environment where there are several persons, we add the work of matching humans across different views, so we finish estimating poses of multi-person. From the experiments, the feasibility and reusability of the proposed method are demonstrated. With this method, the accuracy of results can be maintained while the work is simplified, which can help improve the efficiency of related research.