ARNet: Adaptive Refinement Network Using Context Filter for Motion Forecasting

Motion forecasting is a crucial module for ensuring safety and comfort in autonomous driving. Recently, deep-learning based works usually follow the coarse-to-fine approach, utilizing two-stage models to refine trajectories by fully integrating global context information, such as agent-agent and agent-map interactions. However, these models lack effective strategies for filtering context, which leads to the inclusion of irrelevant information and results in trajectory deviations. To address this issue, a novel two-stage motion forecasting network named AR-Net is proposed in this paper. This method analyzes the trajectory proposals predicted in the first stage and adaptively chooses interactions highly relevant to the agent based on scene differences. Furthermore, the selected context information is aggregated through the global-to-local attention mechanism, enabling the model to accurately capture spatio-temporal dependencies within the scene and thereby improve prediction accuracy. Experiments on the Argoverse motion forecasting dataset reveal that our model achieves promising performance compared with other state-of-the-art approaches.