| Autor: |
Wan, TianJun, Luo, YanMin, Zhang, Zhiqian, Ou, Zhilong |
| Zdroj: |
Signal, Image & Video Processing; Mar2022, Vol. 16 Issue 2, p551-558, 8p |
| Abstrakt: |
Multi-person pose estimation in natural scenes has been a hot topic in the recent years. The prediction speed of the top-down methods is affected by the number of people in the scene, so the bottom-up methods has an advantage in natural scenes. However, the study found that the accuracy of human margin joints (the joints farther from the center of the human, such as wrist and ankle) is always lower than that of the joints that are closer to the center of the human (such as shoulder and hip), and the accuracy gap between joints categories is large. Inspiring from the structural characteristics of human body, this paper proposes a tree structure network (TSNet) for human pose estimation, which divides the joints of the human into several levels according to the characteristics of human body structure, and stepwise predicts the joints from human center to human margin. Combining with the global features, the joint features of the next layer are predicted by extracting the correlation between the joint features of the current layer and the joint features of the previous layer. Therefore, each human joint contains not only the joint information of the current layer and the joint information of the previous layer, but also the background information. The experiment results show that this method can effectively alleviate the uneven precision of joints, and the TSNet can effectively improve the accuracy of lower body joints by setting different activation values for different joints. Extensive experiments on MPII datasets demonstrate the effectiveness of our proposed model and method. [ABSTRACT FROM AUTHOR] |
| Databáze: |
Complementary Index |
| Externí odkaz: |
|
Full text from SpringerLink