动态开放环境下基于交互分析与语义推理的无人车决策技术研究

动态开放环境下存在大量小样本、交互及未知场景。由于小样本事件难建模，交互及未知场景中的交通参与者、道路存在较大不确定性因素，导致无人车决策系统鲁棒性弱、可扩展性差，容易出现决策方法失效和决策结果错误的情况，严重影响无人车在动态开放环境下安全平稳行驶。为此，本项目围绕小样本交互场景下的驾驶意图分类和基于驾驶领域知识表征的决策推理两个科学问题，重点研究：1）基于概率表征的动态车辆交互行为建模，2）小样本情况下的动态车辆意图预测，3）基于语义图谱的自主驾驶决策。以期提出一种能够有效提升动态场景下交通参与者的行为预判准确度的意图预测方法，构建具有高扩展性的驾驶决策语义图谱，并在此基础上，利用虚实结合的实验环境，开展基于驾驶决策图谱的驾驶行为决策方法研究，最终提高动态开放环境下无人车自主决策的安全性和高效性。

There are a large number of few-sample, interactive and unknown scenarios in the dynamic open environment. Since few-sample events are difficult to model, and large uncertainties exist in the interactive and unknown scenarios, the current decision-making systems of autonomous vehicles are weak in robustness and scalability. Once the decision-making system is failed or makes wrong decisions, the safety and smoothness of driving would be strongly affected in the dynamic open environment. Therefore, this project focuses on the following issues: 1) Interactive behavior modeling of dynamic vehicles based on probability representation, 2) Intention prediction of dynamic vehicles based on few-sample learning, 3) Semantic map-based decision-making method. The purpose of this project is to propose an intention prediction method that can effectively improve the accuracy of traffic participants' behavior prediction in dynamic scenarios, and build a highly scalable semantic map for autonomous driving. On this basis, the project will carry out the research of the semantic map-based decision-making method in virtual and real experimental platforms. Eventually, the project is expected to greatly improve the safety and efficiency of decision-making of autonomous vehicles in dynamic open environments.