12 December 2025 in IEEE Internet of Things Journal
Signalized intersections are among the most crash-prone elements of urban traffic networks. Their complex layouts and heterogeneous road users pose substantial challenges for accurate trajectory prediction. Although high-definition (HD) maps are widely used to boost prediction performance, their high construction and maintenance costs and limited update frequency hinder practical deployment. Meanwhile, existing map-free models still struggle to reach the accuracy required in such complex settings. To address these issues, we propose a map-free intersection graph network (MFIG-Net), a map-free, multiagent trajectory prediction framework tailored to signalized intersections. MFIG-Net comprises three components. First, a multisource feature encoder (MSFE) builds per-agent node embeddings by jointly encoding motion histories and static attributes together with traffic-signal phase semantics; the latter are represented using remaining-time sinusoidal positional encodings passed through a multilayer perceptron (MLP) to capture phase transitions. Second, the interaction layer combines a conflict-aware graph attention (CAGA) module—modeling dense spatial interactions via physically grounded, directed edge features and enhanced multihead attention—with a Mamba-based temporal block that captures long-range dependencies through linear-time selective state-space dynamics. Third, a Laplace mixture density network (MDN) decoder produces $K$ multimodal future trajectories per agent with per-timestep scale parameters and mode probabilities, yielding accurate and calibrated uncertainty. Experiments on the SinD dataset from two-phase signalized intersections in Tianjin and Xi’an validate the effectiveness of the proposed model. Additional evaluations on the Argoverse dataset further demonstrate its applicability across diverse urban scenarios, where it outperforms most existing map-free approaches.