With the rapid development of unmanned driving technology, scene understanding and behavior prediction have become one of the core challenges for realizing autonomous driving. This paper proposes a scene understanding and behavior prediction method for unmanned vehicles based on large language model (LLM) and computer vision (CV). By combining the semantic reasoning ability of LLM with the visual perception ability of CV, we designed a more accurate behavior prediction framework that can effectively identify the behavioral intentions of traffic participants in complex traffic environments. Experimental results show that the model based on the combination of LLM and CV surpasses the traditional single CV model in multiple evaluation criteria. In terms of prediction accuracy (mAP), the model reached 91%, which is 8% higher than the traditional CV model; in terms of adaptability in complex scenes, the model showed 89% accuracy, which is 14% higher than the traditional model. In addition, the robustness (AP@0.5) and prediction timeliness (120 ms) of the model in different scenarios also show good performance. Through this study, we have demonstrated the great potential of the combination of LLM and CV in scene understanding and behavior prediction of unmanned driving, and provided new ideas for the development of autonomous driving technology in the future.