Recognition Beyond Perception: Environmental Model Completion by Reasoning for Occluded Vehicles

Martin Krüger, Patrick Palmer, Christopher Diehl, Timo Osterburg, Torsten Bertram

Abstract: It is widely assumed that considering vehicle interactions for trajectory prediction can significantly improve accuracy. All environmental sensors of an automated vehicle (AV) suffer from occlusion. Therefore, relevant vehicles can be occluded by others, especially in dense traffic situations, and remain invisible to the AV. Unobserved vehicles could lead to a drop in trajectory prediction accuracy and even to poor driving behavior. We propose an environmental model completion module that reconstructs unobserved vehicles in occluded areas only using trajectory information of interacting observed vehicles. We demonstrate the functionality of our approach in principle on a toy example and verify the benefit for an AV by improving trajectory prediction results in highly interactive scenes using the proposed module. To the best of our knowledge, we are the first to investigate such a problem on highway data. Reasoning for occluded vehicles is considered more challenging on the highway than in urban scenes since vehicle interactions and cooperation between vehicles are more subtle and nuanced.