In densely urban mixed transportation networks, the dynamic features of the internet of vehicles (IoV) frequently alters connectivity, challenging the safety of connected and autonomous vehicles (CAVs) behavior planning, especially in ramp diverging areas where mixed traffic flows and diverse driving intentions converge. To address this, this paper first construct a multi-view graph neural network (GNN) model that captures spatiotemporal dependencies among CAVs and human-driven vehicles, incorporating speed, intention, and lane-changing behaviors. This multi-perspective model enables a comprehensive understanding of heterogeneous interactions. Building on this, this paper introduce a driving behavior safety prediction mechanism based on artificial potential field (APF) theory to quantitatively evaluate risks between CAVs and surrounding IoV elements, including other vehicles, road boundaries, and destinations. These two components form a collaborative safety behavior planning framework tailored for ramp scenarios, implemented within a graph reinforcement learning framework with a reward function that balances efficiency, safety, and ride comfort. The framework demonstrates strong adaptability to varying traffic densities and driver behaviors. Experimental results on a realistic urban traffic simulation platform demonstrate that compared to benchmarks (A2C, DQN, PPO), the proposed algorithm improves average speed by at least 0.1%, reduces average steps by at least 0.3%, and increases normalized cumulative return by at least 0.5%, while achieving a zero collision rate. These results validate the effectiveness of the proposed framework in enhancing both efficiency and safety in densely urban mixed transportation networks, effectively alleviating congestion and contributing to smoother traffic flow at critical ramp sections, thereby offering a promising solution for future autonomous driving systems.