基于电信数据的室外定位技术研究

越来越多的移动计算依赖位置信息提供基于位置的服务,移动设备的室外定位技术至关重要。目前广为采用的方式是GPS,但移动设备端的GPS位置信息依赖移动设备如手机的GPS传感器获取,电信运营商虽然为用户提供通话和数据服务,却无法获得用户的精确GPS位置。针对这种情况,提出利用手机端和电信基站之间的连接信号数据(简称电信数据),实现移动设备的定位服务。考虑到电信运营商积累了海量的电信数据,因此通过研究基于电信数据的室外定位技术,使得运营商获取用户位置成为可能。提取电信特征数据、以手机所在GPS位置作为标签数据,研究了五种基于机器学习模型的室外定位算法,实现了从基站信号数据到GPS坐标点的预测,通过大量的实验对比了这些方法的定位精度和运行时间、不同数据收集模式的定位精度、不同特征的定位精度以及探索了后处理对定位精度的提升效果。最终通过实验可知,基于栅格化的随机森林分类模型是效果最好的方法,能够达到15~20m的平均误差和10m的中位误差,比前期回归算法在2G和4G数据分别实现了39.46%和54.28%的精度提升,取得与GPS定位接近的定位精度。

Outdoor positioning technology based on telecom data

More and more mobile computing applications rely on location information to provide location based services, and outdoor positioning technology for mobile devices is critical. The widely adopted method is currently GPS, but the GPS position information of the mobile device is dependent on the GPS sensor of the mobile device such as a mobile phone. Although telecom operators provide users with calling and data services, they cannot obtain the users’ precise GPS positions.In view of this situation, we propose to use the connection signal data (telecommunication data for short) between the mobile terminal and the telecom base station to achieve the positioning service of the mobile device. Considering that telecom operators have accumulated vast amounts of telecom data, it is possible to allow operators to acquire user locations by studying outdoor positioning technologies based on telecom data. This paper extracts the telecom feature data, takes the GPS location of the mobile phone as the tag data, and studies five d outdoor positioning algorithm based on machine learning model. The GPS coordinate points are successfully predicted from the base station signal data. A large number of experiments are carried out to compare positioning accuracy,computation time, positioning accuracy of different data collection modes, positioning accuracy of different features among these algorithms. Besides, the effect of post processing on positioning accuracy is explored. Finally, the experiments show that the random forest classification model based on gridding is the best model and can achieve an average error of 15 to 20 meters and a median error of 10 meters. Compared with the previous regression algorithms, the proposal improves the accuracy by 39.46% and 54.28% on the 2G and 4G data respectively, thus achieving the positioning accuracy close to GPS positioning.