Goal oriented trajectory prediction conditioned on reachable road context

Abstract

Trajectory prediction of surrounding traffic agents is crucial for autonomous vehicles to perform collision-free and efficient planning at urban intersections. Despite interactions with neighbour objects, road layout information plays an essential role in improving prediction accuracy and enhancing the interpretability of prediction models. However, exploring reachable areas and effectively leveraging these contextual clues in predictions remain challenging. In this work, a goal-oriented trajectory prediction framework is proposed to integrate valuable road layout information. The framework leverages sparse and non-uniform map elements to represent moving intentions. For effective exploration of relevant map elements, a constrained breadth-first search is proposed, enabling simultaneous and efficient exploration across lateral and longitudinal directions by incorporating behavioural constraints. The attention mechanism and a dynamic mask are combined to focus on the most relevant map element features and predict corresponding goal points, facilitating the final trajectory prediction. This progressive narrowing of the inference space enhances both the accuracy and interpretability of the prediction model. Experimental results on the Intersection Drone Dataset and Roundabout Drone Dataset demonstrate that the proposed model achieves a 73.1% accuracy in predicting the most likely map elements, with an average displacement error below 0.75 m and a final displacement error around 1.95 m with a prediction horizon of 4 s.