In view of the problem found in existing yoga kinesis recognition methods for their failure to mine deep level information such as kinesis and physical features, an improved yoga kinesis recognition algorithm, which is based on multi-feature fusion graph neural network, has thus been proposed. This algorithm utilizes the kinesis record and physique information in the process of yoga, with the advantages of multi-feature fusion and graph neural network combined together. By modeling the relationship between kinesis forms and action, the degree of influence of physique information on different yoga kinesis categories, as well as the short duration and long duration of historical actions, can be obtained. In the experiment, a comparison is made between the performance of the proposed method and other algorithms in yoga kinesis recognition tasks. The results show that the proposed method is characterized with a significant improvement in accuracy, accuracy, recall rate and F1 value, thus verifying the effectiveness of the Yoga kinesis recognition algorithm.