Abstract: Vehicle mobility, resource fluctuations, and regionally distributed non-independent and identically distributed (Non-IID) data jointly compromise training stability and collaborative efficiency in Internet of vehicles (IoV) environments, resulting in model drift, cross-region knowledge conflicts, and long-term convergence instability. Existing studies mainly investigate hierarchical collaboration, personalized modeling, or knowledge distillation independently, whereas the coupled effects of resource dynamics and regional semantic evolution remain insufficiently explored. To address this issue, a resource–semantic collaborative hierarchical federated learning framework is proposed for dynamic regional migration scenarios. The proposed framework adopts a three-tier architecture consisting of clients, edge servers, and a central server. At the client layer, local adaptability is improved by decoupling shared and personalized model components. At the edge layer, adaptive scheduling and regional collaboration are achieved by jointly considering battery level, computational capability, communication bandwidth, and vehicle mobility. At the central server layer, weakly coupled knowledge exchange is performed based on category relevance, thereby reducing semantic interference and communication overhead during cross-region collaboration. Furthermore, theoretical analyses are conducted from the perspectives of training stability, aggregation perturbation, computational complexity, and long-term fairness. Extensive experiments are conducted on an IoV-oriented collaborative visual perception task under independent and identically distributed (IID), Dirichlet, and Non-IID data distributions. Comparisons are made with representative baselines including FedAvg, FedProx, FedDG, and MCA. Experimental results demonstrate that the proposed framework maintains stable convergence under dynamic resource conditions and regional migration. Moreover, the proposed method achieves a balanced performance in terms of model accuracy, training latency, and personalization capability, demonstrating its applicability to continuous collaborative learning in dynamic IoV environments.