基于车载GPS轨迹的立体交叉口空间结构信息获取方法

为了识别立体交叉口中不同的行驶规则, 利用随机森林特征选择方法分析了车辆轨迹数据特征, 按照重要性评分对特征进行聚类; 利用戴维森堡丁指数衡量聚类结果, 获得交叉口最优聚类结果下的各个行驶规则的聚类簇, 并构建聚类簇范围约束的狄洛尼三角网; 利用骨架线提取与公共序列合并方法, 提取立体交叉口的几何结构与拓扑连通关系, 获取城市立体交叉口空间结构信息; 以武汉市2016年出租车轨迹为数据源, 选取了武汉市城区立体交叉口进行空间结构信息获取试验。研究结果表明: 立体交叉口中车载GPS轨迹特征重要性评分的前4项依次是终点角度、起点角度、起终点角度差、中间角度平均值, 其中利用终点角度与起点角度特征组合的聚类结果是最优的; 立体交叉口空间结构信息获取方法在直行、左转、右转方向下识别准确率分别为85.7%、85.4%、87.5%, 综合准确率为86.2%, 直行、左转、右转方向下信息召回率分别为91.5%、87.2%、85.9%, 综合召回率为88.2%, 因此, 较高的准确率与召回率说明本文提出的方法可以准确识别立体交叉口空间结构信息, 并提取立体交叉口中各个行驶规则的几何与拓扑连通关系。 

Information acquisition method of three-dimensional intersection spatial structure based on vehicle GPS trajectory

In order to identify different driving rules at the three-dimensional intersections, the features of vehicle trajectory data were analyzed by using random forest feature selection algorithm, and features were clustered according to the importance scores. The clustered results were measured by Davies-Bouldin index to obtain each driving rule cluster under the optimal clustering result, and Delaunay triangle network was constructed based on the cluster range. The skeleton line extraction and common sequence combination method were used to obtain the geometric structure and topological connectivity relationship of three-dimensional intersection. Finally, the spatial structure information of three-dimensional intersection was obtained. Taking the taxi trajectory data of Wuhan in 2016 as data source, the spatial structure information acquisition experiment of three-dimensional intersection in Wuhan was conduct. Analysis result shows that the top four items of vehicle GPS trajectory feature importance scores are the angle of ending point, the angle of starting point, the difference of starting and ending point angles, and the mean angle of middle points. The clustering result using the characteristics combination of terminal angle and starting angle is optimal. The recognition precision rates of the spatial structure information acquisition method in the directions of straight, left and right turning are 85.7%, 85.4%, and 87.5%, respectively, and the comprehensive precision rate is 86.2%. The information recall rates in the directions of straight, left and right turning are 91.5%, 87.2%, and 85.9%, respectively, and the comprehensive recall rate is 88.2%. The higher precision rates and recall rates indicate that the proposed method can accurately identify the spatial structure information and extract the geometric and topological connectivity relationship of driving rules at three-dimensional intersection.