Heterogeneous Graph Transformer for Connected Autonomous Vehicles' Cooperative Localization

Abstract

Cooperative localization (CL) is a key enabler for future cooperative intelligent transportation system applications. By sharing and fusing multisource information, it can provide high-accuracy localization for connected autonomous vehicles in GNSS-denied or limited environments. Practical CL scenarios often involve heterogeneous cooperative nodes, such as vehicles and landmarks, and relative measurements (e.g., range and bearing). However, existing graph neural network (GNN)-based CL methods do not sufficiently account for scene heterogeneity or fully exploit relative information. Moreover, the state components estimated in CL exhibit substantial scale discrepancies, making effective joint optimization crucial for improving the overall estimation performance of the model. To address these issues, a heterogeneous CL Graph Transformer (HCLGT), together with an adaptive optimization strategy, is proposed for the real-vehicle CL task. First, we incorporate edge-based relative position encoding into the attention computation to better leverage relative measurements. Second, we design relation-specific aggregation modules for different measurement relationships and employ node uncertainty to adaptively adjust attention weights, improving robustness under heterogeneous information quality. Finally, an adaptive loss function is developed to dynamically balance the optimization of multiscale state components. The proposed method is validated on the real-world Leibniz University Cooperative Perception and Urban Navigation Dataset (LUCOOP). Experimental results show that, compared with the best-performing baseline, the proposed method reduces the localization RMSE by 44.61%. Extensive experimental analyses further validate the effectiveness of each module.