MFGAT: Map-Free Trajectory Prediction with Graph Attention Networks for Autonomous Vehicles

Accurate trajectory prediction is a key component for ensuring safe and efficient navigation of autonomous vehicles in complex traffic scenarios. While traditional methods rely heavily on high-definition (HD) maps, these approaches face significant challenges, including high costs, limited availability, and susceptibility to rapid obsolescence. This thesis proposes an end-to-end, map-free trajectory prediction model that leverages Graph Attention Networks (GAT) to dynamically capture spatial-temporal interactions among road agents, eliminating the need for HD maps.The research introduces UNLVTraj, a novel LiDAR-based dataset collected around the University of Nevada, Las Vegas campus, specifically along Cottage Grove Street, Harmon Avenue, and Maryland Parkway. This dataset captures diverse traffic scenarios with annotated trajectories, addressing a gap in existing resources by providing realistic, campus-specific interactions for validation. The proposed model combines GAT with Temporal Convolutional Networks (TCN) and a sequence-to-sequence framework, enabling adaptive weighting of agent interactions to enhance prediction accuracy. Experimental results demonstrate the model’s effectiveness across different environments. On the ApolloScape benchmark, the model achieves a 5.98% reduction in Average Displacement Error (ADE) and a 6.76% reduction in Final Displacement Error (FDE) compared to baseline method. Evaluations on the UNLVTraj dataset further validate its robustness in predicting trajectories for heterogeneous agents, even in unstructured, map-free settings. By offering a scalable, cost-effective solution and a publicly available dataset, this work supports future research in map-free navigation.