Graph Neural Network for Continuous Traffic Density Estimation on Unmonitored Roads from Very Few Scattered Measurements

• A novel architecture is proposed to solve traffic flow reconstruction problems • The architecture leverages graph neural networks and learned feature embeddings • A continuous estimation strategy is proposed to use the solution on real scenarios • The proposed approach is highly generalizable to previously unseen scenarios Traffic control and management play a crucial role within intelligent transportation systems. However, comprehensive traffic monitoring is often constrained by the limited availability of traffic sensors across urban road networks. This paper addresses the problem of Traffic Flow Reconstruction (TFR), which aims to estimate traffic conditions over an entire road network starting from a few sparse traffic sensor measurements. This approach does not require full road coverage by sensors, thereby reducing deployment costs and supporting sustainable urban mobility. In the present research, we propose a lightweight Graph Neural Network (GNN) architecture for real-time TFR using data from a few number of fixed sensors. The model exploits a carefully designed road graph, explicitly augmented to capture traffic propagation and backpropagation effects at junctions, and well-studied node features augmented though a learned encoder. Compared to classical Partial Differential Equations-based TFR solutions, the proposed method achieves satisfactory estimation accuracy while significantly reducing setup times by avoiding scenario-dependent optimizations and precomputation of traffic flow distributions at junctions. Moreover, it overcomes the limitations of existing GNN-based approaches, which often rely on complex architectures that are difficult to adapt to different urban contexts without reprocessing them. Extensive experiments have been conducted and presented in this paper taking into account different moments of the year, various road graph representations and sensors’ positioning, and multiple city environments. Results demonstrate the robustness, generalization capability, and computational efficiency of the proposed approach, making it suitable for large-scale and real-world mobility and transport applications.