A Risk Probability Predictor for Effective Downstream Planning Tasks

Motion prediction predicts the future states of traffic agents, and measures the validity of downstream planning tasks. However, existing predictors are often optimized and evaluated by geometric metrics, without considering the effect of the predictions on planning. The improvement of these metrics alone may not necessarily enhance the performance of the prediction-planning system. In this work, a planning-aware risk probability predictor is proposed that imitates the emphasis human drivers place on traffic agents with reference to their ego plans. Based on a risk-aware decision-making pipeline, we formulate the risk as the product of risk probability and expected collision harm. Interacting with planner, predictor can be informed of future plans and creates a probability field varying with the ego plan. The predictor is evaluated in both handcrafted and recorded real-world scenarios on a test bed with geometric and closed-loop metrics. The findings indicate that assigning varying degrees of significance to traffic agents can assist the planner in making more efficient decisions. Also, the design of predictors should further consider presentations of predictions in combination with the downstream tasks rather than slight improvement in accuracy.