Cooperative perception can effectively address the limitations of individual autonomous vehicles in terms of sensing range and susceptibility to occlusion, making it an indispensable component for the future deployment of autonomous driving. However, different companies equip their autonomous vehicles with various sensors and perception models, which leads to heterogeneous feature maps. To achieve cooperative perception, they must jointly train a feature fusion model or adjust the parameters of pre-trained models, which lacks flexibility and limits the practical application of cooperative perception. To solve this problem, we propose a heterogeneous cooperative perception framework that decouples cooperative perception from single-vehicle perception and offers high flexibility. The framework shares a cooperative perception feature fusion model within the collaborative system. For each type of heterogeneous vehicle joining the collaboration, it fixes the parameters of the pre-trained model and locally trains a feature alignment module to adapt the shared cooperative perception model, achieving the transformation from heterogeneous to homogeneous cooperative perception. By comparing the differences in heterogeneous features, we propose a feature alignment module that integrates both local and global attention mechanisms, effectively converting heterogeneous features to homogeneous ones. Experimental results show that the proposed method achieves an average accuracy of 82% on the heterogeneous cooperative perception datasets OPV2V and OPV2V-H, outperforming eight other cooperative perception methods. The results demonstrate that the proposed method effectively aligns heterogeneous features while maintaining the flexibility of framework.