领域知识与数据驱动的混合交通流车辆轨迹预测

自动驾驶车辆需具备预测周围车辆轨迹的能力。诸多发展中国家普遍存在弱规则、强交互的混合交通流道路,高密度混合交通流的车辆轨迹预测是极具挑战性的任务。为了兼顾混合交通流道路环境下轨迹预测的高精度和可解释性,设计一个融合领域知识和经验的深度学习模型（DK-Conv-LSTM）实现车辆的长、短时轨迹预测。该模型采用卷积结构（Conv）提取交互特征,并将融合车辆历史信息的特征向量送入长短时记忆网络（LSTM）模型实现轨迹预测。知识经验通过嵌入损失函数的方式引导深度学习模型的训练。与基础的LSTM相比,仅添加卷积层结构的Conv-LSTM模型可提升终点轨迹误差（FDE）约30.46 %,提升平均轨迹误差（ADE）约34.78 %;而DK-Conv-LSTM模型可分别提升FDE 46.81 %和ADE 49.08 %;同时DK-Conv-LSTM模型可还原多前车跟驰、超车行为的驾驶轨迹。

Mixed Traffic Flow Trajectory Prediction Method Driven by Domain Knowledge and Data

Autonomous vehicles need to have the ability to predict the trajectory of vehicles around them. There are many mixed traffic flow roads with weak rules and strong interactions in developing countries, and trajectory prediction of high-density traffic flows is an extremely challenging task. In order to predict the trajectory with accuracy and interpretability for mixed traffic flow, a domain knowledge-guided convolutional long short-term memory （DK-Conv-LSTM） to realize the long and short-term trajectory prediction was proposed. In the data-driven model, a convolutional layer （Conv） was used to extract crucial information from interactive vehicles, and the long short-term memory （LSTM） was utilized to predict trajectory after the concatenation of the history information of the vehicle. Knowledge expertise guided the training of deep learning models by being embedded in loss functions. Using the basic LSTM as the benchmark, the Conv-LSTM with only the convolutional structure added, reduces the final displacement error （FDE） by 30.46 % and the average displacement error （ADE） by 34.78 %. The DK-Conv-LSTM reduced the FDE by 46.81 % and ADE by 49.08 %. Moreover, it could recreate complex driving behavior trajectories such as following between two vehicles and overtaking.