一种基于统计推断模型的车辆GNSS轨迹滤波方法

An Algorithm of Filtering Vehicle GNSS Trajectories Based on Statistical Inference

车辆GNSS轨迹是由离散定位点连成的折线段,以往轨迹滤波工作注重离群点及其引起的轨迹线段异常,而忽视了由于采样间隔长而导致的正常点连线严重偏离道路的现象。为解决此问题,该文提出一种基于统计推断的轨迹滤波方法,不依赖路网等辅助数据,在去除离群点的基础上,识别出上述低质量线段。首先采用基于信息熵的最大似然分类法确定线段端点类型,即该端点是位于路口还是道路上;然后再构建规则推断模型,根据线段两端点的空间组合方式,检测出线段是否异常;最后利用北京市2012年11月出租车数据进行验证,采样间隔为50~65s。结果表明,离群点导致的异常线段占32.21%,低采样率导致的亦占6.23%,滤波效果整体良好。

Trajectory filtering is fundamental for spaticr temporal data mining of vehicle GNSS trajec：tories. A GNSS trajectory is discretely represented as a poilyline that is composed of a set of tracking points.A great attention to outliers and related abnor-mal segments has been paid in past wTorks. However, few discussions are found on segments, which are far away from vehicle paths due to a large sampling time interval. In this paper, a statistical interface method was proposed to filter vehicle GNSS traj-ectories. which could detect poor segments from those outliers removed trajectories but did not rely on any auxiliary datasets like road networks. Maximum likelihood classification was firstly used to determine the type of endpoints of a segment based on their information entropy. Rules- based interface model was then built to check the qualification of the segment by concerning the spa-tial pattern of its endpoints.T his method was finally tested using the taxi data collected in Beijing during November,2012, with a sampling interval of 50 to 65 seconds. T he results showed that outliers- introduced from abnormal segments occupied 32 21%, while the ones caused by large sampling time interval were also up to 6. 23%, demonstrating the overall effectiveness of them et ho d.