Autonomous vehicles need to accurately predict the trajectories of surrounding vehicles to ensure safe driving and the safety of other traffic participants. The motion of vehicles in traffic scenario shows multi-modality and is highly influenced by road information and the surrounding vehicles. To address the challenges posed by multimodal properties and social interactions to vehicle trajectory prediction, this paper propose a multimodal trajectory prediction framework called MQNet. The framework incorporates a social interaction module that extracts road semantic information by gridding the road structure. It encodes the trajectory information of neighboring vehicles into a high-dimensional spatial social tensor, capturing the social interaction characteristics between vehicles. Then a motion query-based Transformer decoder with a set of learnable motion queries is proposed to integrate the prior intent of the social interaction module and generate trajectories. The effectiveness of the proposed model is evaluated on the NGSIM US-101 and I-80 datasets. The results suggest that our proposed model outperforms existing baselines in trajectory prediction. Furthermore, a comparative experiment demonstrates that incorporating social interaction during vehicle trajectory prediction can significantly enhance the accuracy of vehicle trajectory prediction.
Motion Query-based Multimodal Vehicle Trajectory Prediction for Autonomous Driving
2023-10-27
3501425 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
MULTIMODAL MOTION PLANNING FRAMEWORK FOR AUTONOMOUS DRIVING VEHICLES
Europäisches Patentamt | 2022
|