Multi-Regulatory Domain OTA Compliance Audit and Semantic Rule Automatic Matching System

Accurate prediction of pedestrian intention and future paths is essential for traffic safety, urban planning, and autonomous navigation. This study develops a multimodal prediction model that combines meaning-based image-text features, motion trajectories, and social interactions. We extract visual-language information from RGB sequences using a CLIP-based encoder and represent group behavior using a Social-GRU network. To improve the reliability of predictions, we apply Bayesian modeling to manage uncertainty. We tested the method on the Waymo and ETH/UCY datasets. On the ETH dataset, the model achieved a 14.2% reduction in average displacement error and a 17.6% reduction in final displacement error, compared with leading baseline methods. The model remained effective in crowded spaces, unclear visual conditions, and sudden motion changes. The results confirm that combining visual-language and motion data improves prediction accuracy. This method offers a practical solution for real-world pedestrian analysis in intelligent transport systems.