Transformer-based multi-task learning for table tennis motion feature recognition
In the process of multi-task sports motion behavior feature recognition, it is prone to be affected by few-shot samples, resulting in catastrophic forgetting phenomena, which leads to poor processing ability of variability. In order to solve the above-mentioned problems, this paper proposes a novel...
| Published in: | Journal of Applied Science and Engineering |
|---|---|
| Main Author: | |
| Format: | Article |
| Language: | English |
| Published: |
Tamkang University Press
2025-06-01
|
| Subjects: | |
| Online Access: | http://jase.tku.edu.tw/articles/jase-202603-29-03-0005 |
| Summary: | In the process of multi-task sports motion behavior feature recognition, it is prone to be affected by few-shot samples, resulting in catastrophic forgetting phenomena, which leads to poor processing ability of variability. In order to solve the above-mentioned problems, this paper proposes a novel table tennis motion feature recognition method based on Transformer-based multi-task learning. This model adopts a grouped attention structure to enhance the extraction ability of local features, and adds the spatial information embedding and temporal information embedding modules to enhance the extraction of spatial and temporal features by the original Transformer model. The extracted chaotic invariant features are classified and recognized through the multi-task learning method by support vector machine to achieve the accurate recognition of multi-task table tennis motion features. The experiment results show that this new method can efficiently identify the motions of table tennis movement, accurately capture the subtle changes of joints, and perform excellently in both single/complex multi-tasks and cross-individual scenarios. |
|---|---|
| ISSN: | 2708-9967 2708-9975 |