Traffic motion forecasting with lightweight network: a novel query sub-attention branch-based cross-attention mechanism

The advancement of autonomous driving requires improved scenario understanding and motion forecasting to ensure safer and more comfortable vehicle journeys. These components are essential for fully replacing human drivers and achieving highly automated driving. In uncertain traffic scenarios, autonomous driving software must anticipate critical situations and act with precision to ensure safety and comfort. Predicting the future trajectories of surrounding agents is vital for autonomous vehicles to operate safely. In this study, we introduce a novel sub-attention module and integrate it with the Cross-Attention mechanism to enhance contextualisation and improve the understanding of semantic relationships among distinct elements within the input sequence, named the Query Sub-Attention branch. Experiments were conducted to validate the proposed mechanism, which was integrated with the Dense-TNT model. The results suggest that the enhanced Cross-Attention mechanism enhances the traffic motion forecasting performance on the Argoverse 2 Motion Forecasting Dataset, showing promise as a generic module for integration with relevant models to advance autonomous driving.