Argoverse 2 study finds 4 urban deceleration modes with early kinematic classifier
Early kinematic signals predict driver deceleration type in 1 second with 75.8% F1 score.
Researchers analyzed 1,219 deceleration events from 234 urban driving logs of the Argoverse 2 Sensor dataset. Using K-means clustering, they identified four stable deceleration modes: anticipatory soft (62.8%), reactive closing (30.6%), brake-like jerk (4.8%), and outlier (1.8%). A HistGradientBoosting classifier using just the first 1.0 second of kinematic data achieved macro-F1=0.758, with scene context contributing +0.059 F1 over kinematics alone. Only pair age shows a medium effect; scene geometry and vulnerable-road-user proximity show negligible effects. The modes are regime-invariant in medium-speed driving (ARI=0.817) but regime-dependent at low speed (ARI=0.166).
- Four deceleration modes identified: anticipatory soft (62.8%), reactive closing (30.6%), brake-like jerk (4.8%), and outlier (1.8%)
- Early-event classifier achieves macro-F1=0.758 using first 1.0 second of kinematic data
- Pair age (time gap) is the only scene context with a medium effect size (epsilon^2=0.085); geometry and VRU proximity are negligible
Why It Matters
Enables autonomous vehicles to predict driver deceleration intent within 1 second, improving safety and ride comfort.