Proper understanding and prediction of driving behavior of surrounding vehicles are one of the most significant requirements for automated driving especially when it comes to safety on a highway. In this paper, we propose a two-layer memory-attention hierarchical model (MAHM) for driving-behavior recognition and motion prediction. This model is based on the human driver’s thinking as well as on brain physiology, i.e., working memory and the selective-attention mechanism. The first layer is a hidden Markov model (HMM), which is used to achieve efficient recognition of driving behavior. The second layer is a memory-attention recurrent neural network (MARNN) for motion prediction, which derives the data from vehicles of interest as input according to driving behavior. Finally, the experimental analysis is performed on the real-data NGSIM US-101 and HighD datasets for highway-driving scenes. We report our results from three perspectives: accuracy of driving-behavior classification, error of predicted trajectories, and execution time.

正确理解和预测周围车辆的驾驶行为是自动驾驶最重要的要求之一，尤其是在高速公路上的安全方面。在本文中，我们提出了一种用于驾驶行为识别和运动预测的两层记忆注意力分层模型（MAHM）。该模型基于人类驾驶员的思维以及大脑生理学，即工作记忆和选择性注意机制。第一层是隐马尔可夫模型（HMM），用于实现对驾驶行为的高效识别。第二层是用于运动预测的记忆注意力循环神经网络 (MARNN)，它根据驾驶行为从感兴趣的车辆中获取数据作为输入。最后，实验分析是在高速公路驾驶场景的真实数据 NGSIM US-101 和 HighD 数据集上进行的。我们从三个角度报告我们的结果：驾驶行为分类的准确性、预测轨迹的错误和执行时间。