Focusing on the issues of Non-Independent and Identically Distributed （Non-IID） data， low-quality data contributed by vehicle participants， and insufficient incentives in Internet of Vehicles （IoV）' Federated Learning （FL）， a quality-aware incentive mechanism based on Stackelberg game for federated learning in IoV was proposed. First， a clustering method based on the similarity of gradient optimization directions of vehicle participants was designed to reduce the impact of Non-IID data on model convergence and improve training efficiency. Second， a dual-dimensional quality evaluation metric was developed that combined gradient direction consistency and loss reduction to measure data quality， and a joint incentive was implemented according to data contribution. Furthermore， a Stackelberg game model was constructed to describe the interaction among cloud servers， edge servers， and vehicle clients， and the optimal strategy combination was derived using backward induction， with the existence and uniqueness of the Stackelberg equilibrium analyzed. Simulation results show that the proposed quality-aware incentive mechanism outperforms baseline mechanisms， achieving significant improvements in both social utility and cloud server utility. Compared with FedAS （Federated parameter-Alignment and client-Synchronization）， the proposed mechanism improves the accuracy by 0.28%， 1.77%， and 5.08% on the MNIST， CIFAR-10， and BelgiumTSC datasets， respectively， verifying the superiority and effectiveness of this mechanism.