ABSTRACT

Lane-changing detection is one of the most challenging tasks in advanced driver assistance system (ADAS). However, modeling driver's lane-changing process is challenging due to the complexity and uncertainty of driving behaviors. To address this issue, a novel sequential model, data-driven lane change detection (DLCD) system is proposed using deep learning techniques. Firstly, DLCD system explores to modeling driving context in spatial domain instead of traditional temporal domain. Secondly, DLCD has an ability of extracting innovative features, i.e. vehicle dynamics feature, lane boundary based distance feature and visual scene-centric feature from multi-modal input data efficiently. Finally, an improved focal loss-based deep long short-term memory (FL-LSTM) network is introduced to learn co-occurrence features and capture the dependencies within lane change events simultaneously. The experimental results on a real-world driving data set show that the DLCD system can learn the latent features of lane change behaviors and significantly outperform other advanced models.

摘要

车道变换检测是高级驾驶辅助系统 (ADAS) 中最具挑战性的任务之一。然而，由于驾驶行为的复杂性和不确定性，对驾驶员的变道过程进行建模具有挑战性。为了解决这个问题，提出了一种使用深度学习技术的新型序列模型、数据驱动的车道变换检测 (DLCD) 系统。首先，DLCD 系统探索在空间域而不是传统的时间域中对驾驶环境进行建模。其次，DLCD 具有从多模态输入数据中高效提取创新特征的能力，即车辆动力学特征、基于车道边界的距离特征和以视觉场景为中心的特征。最后，引入了改进的基于焦点损失的深度长短期记忆 (FL-LSTM) 网络来学习共现特征并同时捕获车道变换事件中的依赖关系。在真实世界驾驶数据集上的实验结果表明，DLCD 系统可以学习变道行为的潜在特征，并显着优于其他先进模型。