**5. Discussion**

With the rapid development of artificial intelligence and its application in various fields, HAR has become an important area of development through deep learning to identify human movement. There is still room for further improvement in the accuracy of current HAR algorithms before its best engineering applications can be achieved.

In the development of existing HAR algorithms, people are always accustomed to introducing newly developed deep learning algorithms into HAR algorithms, which has played a role in improving the accuracy. Compared with traditional machine learning, deep learning essentially uses complex networks for automatic learning data features. In order to achieve better learning of such features, the network of deep learning becomes more and more complex, which requires more expensive hardware, and the requirement is contradictory to engineering application. Therefore, if the network structure remains unchanged (the requirements for hardware also remain unchanged), artificial emphasis on some prior knowledge and enhancement of some features will enable the network to quickly grasp these important features, and improve the accuracy to become a better choice.

Based on the coordination theory in sports kinematics, and by combining the digital robot control theory and the attention mechanism, this study has some innovations in feature enhancement and model structure. For feature extraction, this study uses a twochannel scheme to extract joint and bones features, which are divided into two data streams for analysis. In the aspect of feature enhancement, the coordination attention module and the importance attention module are designed and used to focus on the correlation of upper and lower frames action coordination, and finally achieve the fusion output. This study improves the accuracy, which proves that the idea of HAR combined with the coordination theory is correct.

In addition, we also recognize that because the learning data and validation data of this algorithm come from generally accepted standard datasets, and most of these standard datasets are stable movements of healthy people, this is obviously a positive sample for whole data, and uncoordinated actions should also be the content of learning and analysis, which is one of the defects of this study. Of course, it is easy to imagine that if human movements were inconsistent and the center of gravity was unstable, the predictable results are falls, so this algorithm should be used to predict the action of human falls.
