Vehicle maneuver-based long-term trajectory prediction at intersection crossings

Prediction of surrounding vehicles' turning trajectory and the associated prediction horizon is critical for Autonomous Driving systems navigating urban intersections. In this work, a classical kinematic-based vehicle trajectory, accurate only in short-term and map-agnostic, is fused with a long-term map-aware trajectory derived based on the vehicle's maneuver classification and the associated target lane. The maneuver classification is provided by a Bayesian classifier that incorporates lane-level map information along with past position and kinematics of the vehicle. Although our classifier can recognize multiple types of maneuvers, both the classifier and the proposed trajectory fusion model are here evaluated only with respect to the turn prediction problem, i.e. turning right/turning left versus lane keeping maneuver. Quantitative experiments on simulation datasets of intersection crossing tracks on a typical 4-leg intersection excerpt of a lane-level map show that the proposed approach is effective and robust in extending the trajectory prediction horizon.