With the rapid development of Internet-Connected Vehicles (ICVs), Vehicle-to-Vehicle (V2V) communication has become a key technology for enabling cooperative perception, overcoming the limitations of single-vehicle sensor line-of-sight and occlusion. However, existing methods, while reducing communication overhead, generally ignore the challenges of spatio-temporal constraints and the dynamic optimization problems under long-term communication bandwidth constraints, resulting in insufficient coverage of perception blind spots and a decline in system revenue in high-density dynamic traffic scenarios. To address these issues, this paper proposes a multi-dimensional optimization framework. Firstly, an adaptive vehicle cooperative network is constructed by quantifying four-dimensional features, thereby breaking through the limitations of traditional single-dimensional optimization. Secondly, the Ly-MAPPO algorithm is proposed by combining Lyapunov optimization theory with Multi-Agent Proximal Policy Optimization (MAPPO). The algorithm adopts the centralized training and decentralized execution (CTDE) architecture, uses a virtual bandwidth queue to convert long-term bandwidth constraints into immediate penalty terms, and solves the multi-agent credit allocation problem through counterfactual baselines. Experimental results show that this framework significantly outperforms baseline methods in key metrics such as system revenue, blind zone coverage, average delay and perception accuracy, and reduces communication overhead by 38%.