基于轨迹大数据的立体交叉口精细道路信息获取

As the key link of the urban road network, traffic intersections possess the characteristics of complicated geometrical structures, changeful topological relations, and serious traffic conflicts. At present, the professional approaches including ground data acquisition system, aerial photogrammetry, and space remote sensing, are difficult to meet the demands of rapidly extracting the detailed and up-to-date information of road intersections at a low-cost. In the era of big data, humans as sensors, generate a large number of vehicle trajectories covered each street of cities, which provide a new method for extracting urban road information in terms of both dynamic and static characteristics. The objective of this project is to acquire the detailed geometrical structures and up-to-date topological information of interchanges by using the trace data. Since the positioning precision of trace data is unstable and unreliable, a spatial movement consistency model is proposed to enhance the quality of trajectories. Then, to settle the problems of interchanges recognition including its scale, types, and layers, a multimodal convolutional nerve is applied to accurately identify the interchanges from trace data. Finally, to extract the detailed geometry structures and up-to-date topology information of interchanges from big trace data, a multi-element clustering method is put forward in this project. By these proposed key technologies, this project accomplishes the overstepping of road intersections information acquisition at a large scale and low-cost from the professional mapping to ubiquitous way.

交叉口作为城市路网的关键节点，几何结构复杂、连通关系易变、交通冲突严重，现有地面采集系统、航空摄影测量、航天遥感等手段难以满足精细程度高、现势性好的交叉口信息低成本快速获取需求。大数据时代人人都是传感器，遍布城市大街小巷的轨迹数据，为道路动静态信息获取提供了一种可能。本项目利用轨迹大数据，以立体交叉口精细几何结构与高现势性拓扑信息获取为目标，突破运动一致性的轨迹质量增强技术，解决定位数据不稳定、不可靠难题；构建基于多模态卷积神经网络的立体交叉口识别模型，解决立体交叉口范围、类型、层级识别难题；提出顾及轨迹多要素聚类的立体交叉口精细道路信息提取方法，解决立体交叉口几何结构、拓扑关系的精细提取与现势性检测难题，实现大范围、低成本交叉口信息获取从专业测图到泛在测图的跨越。

交叉口作为城市路网的关键节点，几何结构复杂、连通关系易变、交通冲突严重，现有地面采集系统、航空摄影测量、航天遥感等手段难以满足精细程度高、现势性好的交叉口信息低成本、快速获取需求。大数据时代下对地观测技术从静态到动态，从专业到泛在，从人工到智能，“人人都是传感器”，遍布城市大街小巷的车载轨迹，为道路动/静态信息获取提供了一种新的技术可能。本项目利用众源轨迹大数据，以立体交叉口精细几何结构与高现势性拓扑信息获取为目标，提出融合小波和相似度聚类的三维轨迹数据增强算法，实现轨迹数据平面几何精度、高程数据异常值去除与位置偏差改善；针对立体交叉口识别困难问题，提出基于Mask-RCNN的城市立体交叉口精准识别方法，突破城市大区域复杂立体交叉口识别、分类、范围获取；针对立体交叉口几何结构各异、连通关系复杂，研究立体交叉口轨迹转弯规律与高程变化特点，构建顾及轨迹方向、平面位置、高度多要素增量融合与语义分割算法，实现立体交叉口三维几何结构，如：路段中心线、立体分层信息、匝道出入口、坡道起止点，及拓扑关系的精细获取。实验以武汉市、上海市、北京市、罗马等城市的众源轨迹数据为例，对轨迹数据质量增强、立体交叉口识别、立体交叉口三维几何、拓扑信息获取进行了验证与分析。实验结果表明，众源轨迹数据按照本项目所提方法进行增强处理后，高程数据与DEM数据之间的偏差从平均值12.89m、标准差51.4m提升至平均值10.67m、标准差8.6m，平面数据定位精度则可以按需实现分米级、米级精度的滤选；采用Mask-RCNN模型的交叉口识别精度可以达到95%、召回率为91%；立体交叉口道路中心线获取准确率为94%、匝道分离（汇入）点提取准确率为89%、道路坡道起止点准确率为95%、立体分层信息获取准确率为96%。实验验证本项目所提方法可以实现大范围、低成本交叉口信息获取从专业测绘到泛在测图的跨越，为构建智慧城市提供基础数据支撑。