基于GCN-Transformer的车辆换道行为建模与轨迹预测方法

A Vehicle Lane-Changing Behavior Modeling and Trajectory Prediction Method Based on GCN-Transformer

摘要: 对车辆换道行为建模并准确预测未来行驶轨迹对交通流的稳定与安全至关重要，为了解决目前大多数轨迹预测模型在同时捕捉车辆之间的空间相关性和时间依赖性上能力不足的问题，结合车辆轨迹的时空特点，本研究提出了一种基于长短期记忆网络、图卷积网络和Transformer编码器的改进建模策略。首先利用长短期记忆网络，对目标车辆和周围车辆在换道临界点前三秒内的状态信息分别进行轨迹编码，接着通过图卷积网络提取空间交互特征，然后通过Transformer编码器提取时间交互特征，最后将前三个模块处理后的特征向量合并后，输入至长短期记忆网络进行解码，得到目标车辆未来五秒的行驶路径预测输出。在NGSIM数据集和HighD数据集上进行实验，并与多种基准模型对比，结果表明：在2秒内的预测时域下，本文模型与PiP模型和DLM模型不差上下，但优于其他LSTM改进模型；在3~5秒内的预测时域下，本文模型优于各基准模型。本文还通过消融实验，证明了设计的时空特征提取模型对模型准确预测的有效贡献。

Abstract: Modeling vehicle lane change behavior and accurately predicting future driving trajectory is crucial for the stability and safety of traffic flow. In order to solve the problem that most current trajectory prediction models are insufficient in capturing the spatial correlation and time dependence between vehicles at the same time, combined with the spatio-temporal characteristics of vehicle trajectory, this study proposes an improved modeling strategy based on long short-term memory network, graph convolutional network and Transformer encoder. Firstly, the long short-term memory network is used to encode the state information of the target vehicle and the surrounding vehicles within three seconds before the lane change critical point. Then, the spatial interaction features are extracted through the graph convolution network, and then the time interaction features are extracted through the Transformer encoder. Finally, the feature vectors processed by the first three modules are merged and input into the long short-term memory network for decoding, and the driving path prediction output of the target vehicle in the next five seconds is obtained. Experiments are conducted on the NGSIM dataset and HighD dataset, and compared with a variety of benchmark models. The results show that: in the prediction time domain of 2 seconds, the proposed model is similar to the PiP model and DLM model, but better than other LSTM improvement models; in the prediction time domain of 3~5 seconds, the proposed model is superior to the benchmark models. This paper also demonstrates the effective contribution of the designed spatiotemporal feature extraction model to the accurate prediction of the model through ablation experiments.