To address the issues of scarce annotated samples and insufficient feature representation capability in sign language recognition tasks, this study proposes a pre-training method for sign language visual features based on self-supervised learning. By constructing a spatial-temporal joint augmentation mechanism and a contrastive learning framework, spatiotemporal feature representation is optimized under unsupervised conditions. A dual-branch encoding architecture integrating C3D and Transformer is designed to balance local motion dynamics and global semantic dependencies. An occlusion reconstruction auxiliary task is introduced to enhance the model’s ability to recover key motion regions and missing information. After pre-training, a frozen transfer strategy is adopted to migrate features to downstream sign language recognition tasks, combined with BiLSTM-based temporal modeling and CTC-based classification decoding to build the overall system. Experimental results on the RWTH-PHOENIX-Weather 2014T dataset demonstrate a recognition accuracy of 85.7% and a character error rate reduced to 10.3%, significantly outperforming traditional transfer learning approaches and validating the effectiveness and generalization advantages of the proposed self-supervised pre-training mechanism in low-annotation resource scenarios.