Driving maneuvers prediction is one of the most challenging tasks in modern Advanced Driver Assistance System. Such predictions can improve driving safety by alerting the driver to the danger of unsafe or risky traffic situations. In this research, we presents a novel Attention-based Global Context Network (AGCNet) for driving maneuvers prediction from multiple modality data, including front view video data and driver physiological signals. Firstly, with Global Context block, the AGCNet has an ability of modeling the long-range dependency contextual features from multi-modal data with lightweight computation. Secondly, the Channel-wise Attention is introduced in AGCNet to focus on valuable features. Finally, a custom-built Dual attention-based Long Short-Term Memory (DaLSTM) network is designed to learn co-occurrence features and predict driving maneuvers. Specially, the DaLSTM can employ attention mechanisms over heterogeneous features and time steps simultaneously. The experimental results show that the AGCNet is capable of learning the latent features of driving maneuvers and achieving significantly better performance than other advanced models on a real-world driving dataset.

驾驶动作预测是现代高级驾驶辅助系统中最具挑战性的任务之一。这种预测可以通过提醒驾驶员注意不安全或危险交通情况的危险来提高驾驶安全性。在这项研究中，我们提出了一种新颖的基于注意力的全局上下文网络（AGCNet），用于从多种模态数据（包括前视视频数据和驾驶员生理信号）中进行驾驶机动预测。首先，借助 Global Context 块，AGCNet 能够通过轻量级计算从多模态数据中对远程依赖上下文特征进行建模。其次，在 AGCNet 中引入 Channel-wise Attention 以专注于有价值的特征。最后，一个定制的基于双重注意力的长短期记忆 (DaLSTM) 网络旨在学习共现特征和预测驾驶操作。特别是，DaLSTM 可以同时在异构特征和时间步上采用注意力机制。实验结果表明，AGCNet 能够学习驾驶动作的潜在特征，并在真实世界驾驶数据集上实现比其他高级模型更好的性能。