As the Internet of Everything becomes the trend of the times, traditional video coding and compression methods are difficult to remove a large amount of redundant information in video data, which will inevitably reduce transmission efficiency. To address this challenge, a semantic communication-oriented 3D skeleton data source encoding and compression method (DMDCT) is proposed. For the redundancy problem in the skeleton data, starting from the semantic concept, a multi-scale skeleton representation method is proposed, which adaptively describes the motion state of skeleton participating in each different action semantics while retaining the human skeleton structure. Discrete Cosine Transform (DCT) separates the DC and AC components represented by multi-scale skeleton points from the frequency domain level, further reducing the overall data volume. Different from the traditional communication method of transmitting original video data, combined with semantic communication, only skeleton point data related to high-level tasks is transmitted, which improves the data transmission efficiency. Experiments on the public dataset NTU RGB+D taking action recognition as an example show that, under the same compression rate, DMDCT's TOP-1 accuracy rate is about 5% higher than that of similar algorithms, and retaining only 10% of the DCT coefficients can still achieve an accuracy of 74.2%, while the data volume is only 6% of the original data volume.