The wide proliferation of sensors and devices of Internet of Things(IoT), together with Artificial Intelligence (AI), has created the so-called Smart Environments. From a network perspective, these solutions suffer from high latency and increased data transmission. This paper proposes a Federated Learning (FL) architecture for Real-Time Traffic Estimation, supported by Roadside Units (RSU’s) for model aggregation. The solution envisages that learning will be done on clients with their local data, and fully distributed on the Edge, with high learning rates, low latency, and less bandwidth usage. To achieve that,this paper discusses tools and requirements for FL implementation towards a model for real-time traffic estimation, as well as how such solution could be evaluated using VANET and network simulators. As a first practical step, we show a preliminary evaluation of a learning model using a data set of cars that demonstrate a distributed learning strategy. In the future, we will use a similar distributed strategy within our proposed architecture.