Abstract
Accurate prediction of agent motion trajectories is crucial for autonomous driving, contributing to the reduction of collision risks in human-vehicle interactions and ensuring ample response time for other traffic participants. Current research predominantly focuses on traditional deep learning methods, including convolutional neural networks (CNNs) and recurrent neural networks (RNNs). These methods leverage relative distances to forecast the motion trajectories of a single class of agents. However, in complex traffic scenarios, the motion patterns of various types of traffic participants exhibit inherent randomness and uncertainty. Relying solely on relative distances may not adequately capture the nuanced interaction patterns between different classes of road users. In this paper, we propose a novel multi-class trajectory prediction method named the social force embedded mixed graph convolutional network (SFEM-GCN). SFEM-GCN comprises three graph topologies: the semantic graph (SG), position graph (PG), and velocity graph (VG). These graphs encode various of social force relationships among different classes of agents in complex scenes. Specifically, SG utilizes one-hot encoding of agent-class information to guide the construction of graph adjacency matrices based on semantic information. PG and VG create adjacency matrices to capture motion interaction relationships between different classes agents. These graph structures are then integrated into a mixed graph, where learning is conducted using a spatiotemporal graph convolutional neural network (ST-GCNN). To further enhance prediction performance, we adopt temporal convolutional networks (TCNs) to generate the predicted trajectory with fewer parameters. Experimental results on publicly available datasets demonstrate that SFEM-GCN surpasses state-of-the-art methods in terms of accuracy and robustness.
Abstract (translated)
准确预测代理的运动轨迹对自动驾驶至关重要，有助于减少人与车辆互动中的碰撞风险，并为其他交通参与者确保充足的反应时间。目前的研究主要集中在传统深度学习方法，包括卷积神经网络（CNNs）和循环神经网络（RNNs）。这些方法利用相对距离预测单一类代理的运动轨迹。然而，在复杂的交通场景中，不同类型交通参与者的运动模式表现出固有的随机性和不确定性。仅依赖相对距离可能不足以捕捉不同类别道路用户之间的细微交互模式。在本文中，我们提出了名为社会力嵌入混合图卷积神经网络（SFEM-GCN）的新颖多类轨迹预测方法。SFEM-GCN由三个图结构组成：语义图（SG）、位置图（PG）和速度图（VG）。这些图编码了复杂场景中不同类别代理之间的社会力关系。具体来说，SG利用代理类信息的one-hot编码来引导构建基于语义信息的图邻接矩阵。PG和VG创建邻接矩阵以捕捉不同类别代理之间的运动交互关系。然后将这些图结构整合成一个混合图，使用时空图卷积神经网络（ST-GCNN）进行学习。为了进一步提高预测性能，我们采用时间卷积网络（TCNs）生成预测轨迹，同时参数更少。公开可用数据集上的实验结果表明，SFEM-GCN在准确性和鲁棒性方面超过了最先进的 methods。
URL
https://arxiv.org/abs/2404.13378
