Weighted Three-Stream Spatial-Temporal Graph Convolutional Network Based on Novel Adaptive Skeleton Features
Uloženo v:
| Název: | Weighted Three-Stream Spatial-Temporal Graph Convolutional Network Based on Novel Adaptive Skeleton Features |
|---|---|
| Autoři: | Pei Jianfeng, Bi Sheng |
| Informace o vydavateli: | Springer Science and Business Media LLC, 2023. |
| Rok vydání: | 2023 |
| Popis: | Graph convolution-based methods have achieved remarkable performance in skeleton-based action recognition. However, the mainstream graph convolution methods all use the spatial position coordinates of the skeleton joints as the feature inputs, and less research has been done on mining the second-order semantic information of the skeleton features. The mainstream features use a uniform standard for all the motion objects. In this paper, we propose a new adaptive skeleton feature extraction algorithm, that can adaptively determine the action center of the motion according to different motion objects, different video frames, and different positional changes. It can generate unique skeleton features belonging to each video. Our features are more flexible than mainstream feature selection, which is centered around the origin. Extensive experiments on the mainstream NTU-RGB-D dataset show that the proposed feature extraction algorithm significantly improves recognition accuracy in single-stream and multi-stream frameworks. |
| Druh dokumentu: | Article |
| DOI: | 10.21203/rs.3.rs-3368490/v1 |
| Rights: | CC BY |
| Přístupové číslo: | edsair.doi...........ce32c85bb941f8e7aaf327ab3c5dc288 |
| Databáze: | OpenAIRE |
Nájsť tento článok vo Web of Science | Abstrakt: | Graph convolution-based methods have achieved remarkable performance in skeleton-based action recognition. However, the mainstream graph convolution methods all use the spatial position coordinates of the skeleton joints as the feature inputs, and less research has been done on mining the second-order semantic information of the skeleton features. The mainstream features use a uniform standard for all the motion objects. In this paper, we propose a new adaptive skeleton feature extraction algorithm, that can adaptively determine the action center of the motion according to different motion objects, different video frames, and different positional changes. It can generate unique skeleton features belonging to each video. Our features are more flexible than mainstream feature selection, which is centered around the origin. Extensive experiments on the mainstream NTU-RGB-D dataset show that the proposed feature extraction algorithm significantly improves recognition accuracy in single-stream and multi-stream frameworks. |
|---|---|
| DOI: | 10.21203/rs.3.rs-3368490/v1 |