Driver lane change intention (DLCI) predicting has become an essential research for the development of human–machine co-driving system. This work makes an attempt to predict the DLCI, which is the result of complex interaction between human drivers and driving scene. While few works have explored the relationship between driver behavior features and key features of driving scene when predicting the DLCI. To solve this gap, we developed a DLCI prediction method based on topological graph constructed by driver behaviors and traffic context. However, challenges trend on the heels of that because of some unavoidable features that are irrelevant to the DLCI prediction in the topological graph, the difficulty of capturing global dependencies in the driver’s head pose sequence, the dynamics of the relationship between different categories, and insufficient of the DLCI dataset. Therefore, we designed a DLCI predicting model based on dynamic graph convolution network with semantic attention module (DGCN-SAM) and self-supervised guided learning based on the understanding of topological graph (SGL-UTG). Specifically, an invert residual module with anthropomorphic attention mechanism (IRM-AAM) was designed to extract important features in topological graphs. The Transformer with multi-head self-attention was used to capture the global dependences of driver’s head pose sequence. DGCN-SAM was developed to model the relationship between different categories or nodes in the graph. And SGL-UTG was proposed to improve the generalization performance and prevent overfitting in the absence of sufficient DLCI data. The experimental results demonstrate that the proposed method can predict the DLCI in real-time with high accuracy.

中文翻译：

---

基于驾驶员行为和交通情境构建拓扑图的人机协同驾驶系统驾驶员换道意图预测

驾驶员变道意图（DLCI）预测已成为人机协同驾驶系统开发的重要研究。这项工作尝试预测 DLCI，这是人类驾驶员与驾驶场景之间复杂交互的结果。而很少有研究在预测 DLCI 时探讨驾驶员行为特征与驾驶场景关键特征之间的关系。为了解决这一差距，我们开发了一种基于驾驶员行为和交通环境构建的拓扑图的 DLCI 预测方法。然而，由于拓扑图中一些不可避免的与 DLCI 预测无关的特征、捕获驾驶员头部姿势序列中的全局依赖关系的困难、不同类别之间关系的动态性以及不足，挑战趋势随之而来。 DLCI 数据集。因此，我们设计了一种基于带有语义注意模块的动态图卷积网络（DGCN-SAM）和基于拓扑图理解的自监督引导学习（SGL-UTG）的DLCI预测模型。具体来说，设计了具有拟人注意机制的逆残差模块（IRM-AAM）来提取拓扑图中的重要特征。具有多头自注意力的 Transformer 用于捕获驾驶员头部姿势序列的全局依赖性。DGCN-SAM 的开发是为了对图中不同类别或节点之间的关系进行建模。提出SGL-UTG是为了提高泛化性能并防止在没有足够DLCI数据的情况下过拟合。实验结果表明，该方法能够实时、准确地预测DLCI。