基于多通道态势图的自动驾驶场景表征方法

开展自动驾驶测试场景研究能够大幅减少自动驾驶汽车的测试周期与开发成本，是未来评价和提升自动驾驶技术的重要基础。为此，联合基于本体论的场景解构方法，提出了一种基于多通道态势图的自动驾驶场景表征方法，并对多通道态势图的场景聚类与场景复杂度进行研究。首先，对目前的自动驾驶测试方法进行分析，论述道路测试的不足之处以及基于场景的自动驾驶虚拟测试的优点，并对当前的场景解构与表征方法进行了总结；然后，运用本体论解构场景中的信息，并建立场景的本体模型，对模型中的数据属性进行参数化；接着，对真实场景、场景中的语义信息和多通道态势图场景进行对比分析，定义表征场景的多通道态势图的数据格式，将解构出的场景信息重组到多通道态势图的不同层中；之后，以汉明距离为基础设计了多通道态势图的对象层相似度计算方法，采用K均值聚类算法对驾驶场景对象层进行聚类分析，并借助层次分析法对基于多通道态势图的驾驶场景复杂度计算进行研究；最后，以KITTI数据集的一些真实场景为例，绘制场景开始时刻的多通道态势图，分析聚类出的9种对象分布类型。研究结果验证了多通道态势图场景复杂度计算方法的有效性。

Automatic Driving Scenario Representation Based on Multi-channel Situation Map

Research on autonomous driving scenarios can greatly reduce testing cycle duration and development cost, which are important factors in the evaluation and improvement of future autonomous driving technologies. Therefore, an automatic driving scenario reconstruction method using a multi-channel situation map was proposed based on an ontology-based scenario deconstruction method. The scenario clustering and scenario complexity of the multi-channel situation map was studied. First, the current automated driving test methods were analyzed. Subsequently, the shortcomings of road testing and advantages of scenario-based automated driving virtual testing were summarized. Then, the current scenario deconstruction and reconstruction methods were reviewed. Ontology was used to deconstruct the information in the scenario, and an ontology model was established to parameterize the properties of data in the model. The real scenario, semantic information in the scenario, and multi-channel situation map were compared and analyzed. The format of the multi-channel situation map was defined, and the deconstructed scenario information was reconstructed into different layers of the multi-channel situation map. In the object layer clustering research, a K-means clustering algorithm was used, and a new method to calculate object layer similarity was designed based on the Hamming Distance. Subsequently, an analytic hierarchy process was used to calculate the complexity of the multi-channel situation map. Finally, some real scenarios from the KITTI data set were considered for scenario deconstruction and reconstruction, and a multi-channel situation map was created at the beginning of the scenario. From the clustering, a total of nine object distribution types were analyzed. The research result verifies the effectiveness of the scenario complexity calculation method.