Given semantic information loss in multimodal data transmission and 3D reconstruction accuracy degradation of human bodies caused by channel noise interference, this study proposes a 3D reconstruction method for human bodies based on self-supervised multimodal semantic communication (SMH3D). This method utilizes cross-modal contrastive self-supervised learning to achieve feature alignment of multimodal data, such as RGB images and depth maps, and compresses the features by a variational self-encoder. Then, a knowledge graph-enhanced semantic coding mechanism and a variable-length coding strategy are introduced to prioritize the protection of key semantic information under limited bandwidth. To solve the instability of real channel conditions, this study designs the end-to-end channel coding and decoding modules and employs channel state feedback to adaptively adjust the transmission parameters, thus ensuring the high-fidelity transmission of semantic features in harsh environments such as noise and fading. A multimodal feature fusion strategy based on the combination of Transformer and graph neural networks, and the conditional implicit function is adopted to realize high-accuracy 3D model reconstruction of human bodies. The experimental results show that SMH3D significantly outperforms the traditional point cloud transmission method SemCom and the semantic communication method SCS for real-time 3D reconstruction tasks in terms of IoU, PSNR, and semantic fidelity, showing stronger robustness and stability in low signal-to-noise ratio environments.