Abstract: The traditional wireless resource allocation methods for internet of vehicles in highly dynamic change scenarios cannot meet the high-quality communication demand, especially in congested locations such as intersections, where there are problems such as inefficient resource scheduling, network congestion, and high time budget. To address the above problems, a multi-intelligence reinforcement learning strategy based on the attention mechanism is proposed, aiming to enhance the model's ability to reasonably schedule and classify resources for multiple intelligence by improving the design of the multi-intelligence training model. The paper introduces an attention-based mechanism to learn the dependency relationship between multiple intelligence, constructs global state semantic features with multiple perspectives, and enhances the strategy evaluation ability of global value function during the training process of multi-intelligence reinforcement learning strategy. Simulation validation shows that the proposed method effectively enhances the resource allocation capability of the policy in the spectrum sharing task and improves the delivery success rate of the total channel capacity and the payload of the inter-vehicle communication link.