The paper introduces a novel method for generating optimal trajectories by combining predictions from an ensemble of heterogeneous trajectory prediction models. The key insights are:
Sampling based solely on predicted probabilities (e.g., selecting the top-k most likely trajectories) degrades performance as more ensemble members are added, due to misalignment between model outputs.
The proposed method frames trajectory sampling as a risk minimization problem, where the ensemble approximates the true risk. This allows the method to effectively leverage the diversity of the ensemble to generate a set of optimal trajectories.
Extensive experiments on the nuScenes dataset demonstrate that the proposed method outperforms existing sampling techniques, including alternatives like K-Means and Non-Maximum Suppression. It also surpasses the performance of the individual base learners (PGP, LaPred, LAformer) by a large margin.
The paper provides a comprehensive empirical study on the influence of ensemble composition, showing that a mixed ensemble of strong models (e.g., LAformer and PGP) performs best, while dropout-based ensembling is less effective.
Overall, the paper presents a novel and effective approach for trajectory prediction in autonomous driving, highlighting the potential of advanced ensembling techniques to significantly improve predictive performance.
A otro idioma
del contenido fuente
arxiv.org
Consultas más profundas