How do self-driving cars predict danger before it happens?

Autonomous vehicles need to predict where other cars will go—but most trajectory models ignore how danger evolves ahead of time. This work adds a "risk horizon profiling" layer that tracks spatial-temporal proximity of nearby objects across future time steps, letting the model learn which moments matter most to human drivers. Tested on highway and urban crash data, the method cuts prediction error by 25–29% over baselines and generalizes across safe, near-crash, and actual crash scenarios. Code released.