Federated learning allows multiple users and parties to collaborate and train machine learning models in a distributed and privacy-preserving manner in Vehicular Adhoc Networks VANETs. This computing paradigm addresses privacy concerns; however, it comes at a considerable cost of network resources. After training the machine learning models in conventional federated learning frameworks, devices share that model with a central server, mostly cloud, where the global aggregation is performed. Multiple devices communicating with a central server raise network bandwidth and congestion concerns. To solve this problem, we proposed a federated learning framework for VANETs where instead of using a fixed global aggregator, we used variable global aggregation nodes. The global aggregation node is selected based on communication delay and workload in the proposed framework. We also believe that, in a vehicular Adhoc network, all network nodes cannot participate in the learning process due to network, computation, and energy resource limitations. We Also proposed a client selection algorithm that adapts itself and selects some clients based on specific criteria. Finally, the proposed technique is compared with the hierarchical federated learning framework (HFL) and FedAvg where proposed method outperformed in terms of accuracy.