[论文解读] Transferable and Adaptable Driving Behavior Prediction

Letian Wang, Yeping Hu+4 more|arXiv (Cornell University)|Feb 10, 2022

Autonomous Vehicle Technology and Safety被引用 20

一句话总结

提出 HATN，一种分层、可迁移且可适应的驱动行为预测框架，结合高层语义图驱动的意图模型与低层轨迹生成器以及在线自适应模块。

ABSTRACT

While autonomous vehicles still struggle to solve challenging situations during on-road driving, humans have long mastered the essence of driving with efficient, transferable, and adaptable driving capability. By mimicking humans' cognition model and semantic understanding during driving, we propose HATN, a hierarchical framework to generate high-quality, transferable, and adaptable predictions for driving behaviors in multi-agent dense-traffic environments. Our hierarchical method consists of a high-level intention identification policy and a low-level trajectory generation policy. We introduce a novel semantic sub-task definition and generic state representation for each sub-task. With these techniques, the hierarchical framework is transferable across different driving scenarios. Besides, our model is able to capture variations of driving behaviors among individuals and scenarios by an online adaptation module. We demonstrate our algorithms in the task of trajectory prediction for real traffic data at intersections and roundabouts from the INTERACTION dataset. Through extensive numerical studies, it is evident that our method significantly outperformed other methods in terms of prediction accuracy, transferability, and adaptability. Pushing the state-of-the-art performance by a considerable margin, we also provide a cognitive view of understanding the driving behavior behind such improvement. We highlight that in the future, more research attention and effort are deserved for transferability and adaptability. It is not only due to the promising performance elevation of prediction and planning algorithms, but more fundamentally, they are crucial for the scalable and general deployment of autonomous vehicles.

研究动机与目标

• 让人类式的分层认知（高层插槽插入与低层轨迹跟踪）成为动机并建模，以提高在不同场景中的预测可迁移性。
• 开发一个紧凑、通用的表示（Semantic Graph with Dynamic Insertion Areas）以支持场景迁移。
• 结合在线自适应以捕捉个体和场景特定行为变化。
• 在真实交通数据（INTERACTION 数据集）上展示更高的预测准确性、迁移性和适应性。

提出的方法

• 引入 HATN（Hierarchical Adaptable and Transferable Network），包含三个组成部分：用于意图推理的高层 Semantic Graph Network (SGN)，用于轨迹生成的低层 Encoder-Decoder Network (EDN)，以及使用修改后的 Extended Kalman Filter (MEKFλ) 的在线自适应 (OA) 模块。
• 将驾驶场景表示为 Semantic Graphs，其中 Dynamic Insertion Areas (DIA) 作为图节点，引导意图和目标状态分布。
• SGN 通过高斯混合模型输出 DIAs 的插入概率 w_t 和目标状态分布 g_t。
• EDN 生成未来轨迹，条件是历史动态 S_{t-T_h:t} 与意图信号 g_t。
• OA 在线适应 EDN 参数 θ 以使用 MEKFλ 最小化预测误差。
• 将高层问题表述为预测 Y_{t+1:t+T_f} = f_HATN(O_{t-T_h:t})。

实验结果

研究问题

• RQ1 分层、语义基础的表示是否能够在不同驾驶场景（交叉口、环岛等）之间实现零样本迁移？
• RQ2 在线自适应是否提升驾驶行为预测的个性化和场景迁移能力？
• RQ3 带有 Dynamic Insertion Areas 的语义图是否能有效捕捉包含丰富交互的驾驶情境，用于意图与轨迹预测？
• RQ4 在真实世界数据上，与最先进方法在预测准确性、可迁移性和适应性方面的比较如何？

主要发现

• 通过将高层意图与低层轨迹生成分离，HATN 框架实现了更高的预测质量。
• 带 Dynamic Insertion Areas 的 Semantic Graph 提供了紧凑、场景无关的表示，支持跨场景迁移。
• 通过 MEKFλ 的在线自适应通过在线更新 EDN 参数，提高对个体和场景的定制化。
• 在 INTERACTION 数据集上的实验在准确性、可迁移性和适应性方面优于基线（并有大量消融实验）。
• 本文提供了一个认知层面的解释，说明为什么分层、可迁移且可适应的设计能够在密集交通中获得更好的驾驶行为预测。