Vehicle group-based multivehicle spatiotemporal interaction pattern learning

Interpreting naturalistic multi-vehicle interactions is vital for human-like autonomy and interactive testing. We model interactions from a group perspective, a common yet understudied phenomenon. A framework is proposed for learning spatiotemporal interaction patterns from vehicle groups. A Multi-Vehicle Interaction Directed Network (MIDN) is introduced and constructed from interaction strength (IS) and spatial relations, with influences from non-neighbouring vehicles incorporated. Patterns are analysed with an Attention Auto-encoder and k-means. Using naturalistic trajectory data, the approach is validated: groups are identified with high within-group and low between-group IS, and chained groups are found to be more prevalent than multi-lane groups. Six interpretable patterns are identified - Increasing/Decreasing Backward (IBT/DBT), Increasing Forward (IFT), Rear Decreasing (RDT), Overall Maintenance (OMT), and Multiple Fluctuations (MFT) - and are shown to correlate with speed and acceleration profiles. The findings are expected to support human-like planning in autonomous vehicles, the construction of test scenarios, and roadside traffic safety monitoring.