ABSTRACT
This paper introduces LGTraj, a diffusion-based framework for long-horizon trajectory prediction in urban driving scenarios, addressing challenges like mode collapse, computational inefficiency, and traffic constraint incorporation.
PAPER · PDF
Loading PDF...
Key findings
Proposes Physics-Aware Classifier-Free Guidance to enforce kinematic constraints.
Introduces Social Interaction Guidance for modeling multi-agent dependencies.
Develops Temporal Coherence Guidance for consistency across long prediction horizons.
Demonstrates state-of-the-art performance on nuScenes and Argoverse 2 benchmarks.
Limitations & open questions
The paper does not discuss the computational resource requirements for real-time applications.
Further analysis on the generalizability of the model to different urban environments is needed.