The perception quality and energy consumption of millimeter-wave radars for autonomous vehicles are difficult to balance in dynamic road environments. In order to address this issue, an active cooperative scheme based on perception signal-to-noise ratio (SNR) prediction and multi-agent reinforcement learning was proposed to investigate the vehicle task allocation. First, a spatiotemporal generation model of dynamic SNR in a multi-vehicle environment was constructed. Subsequently, a CGAN-LSTM framework was used to predict the average perceived SNR of future time slots. Moreover, an improved CGAN network was embedded into the observation module of the agents. Combined with the distributed algorithm framework of Recurrent Multi-Agent Proximal Policy Optimization (rMAPPO), vehicle task allocation and power control under power limitations were optimized. The results show that, compared with traditional methods, the SNR prediction accuracy of the vehicle millimeter-wave radar is significantly improved. Under the same energy consumption constraint, the average perceived SNR of the system is increased by 21%, and more efficient sensing coverage is ensured. It is concluded that the proposed active cooperative scheme effectively enhances the perception performance and energy efficiency of autonomous vehicles in dynamic environments.