Cell Tower Based High Performance Outdoor Localization Methods

It is still a hotspot for researchers to locate mobile devices using measurement report(MR)data.Localization can be divided into two categories for different locating areas,which are indoor localization and outdoor localization.The locating area of outdoor localization is much larger than indoor localization,so it usually harder to design high performance outdoor localization algorithms.In this paper,we focus on cell-tower based outdoor localization methods.Our work is mainly from two aspects.First,we design a method to auto-label measurement report data.Second,we propose two contextual localization algorithms.We use one-month data of 17699 users from one telco operator in Shanghai,which covers 13 square kilometers in centroid of the city.There are nearly 100 million training samples after labeling,and we split the dataset into training set and testing set.The main contribution of our work is listed as follows:(1)Car drive is often used to collect labeled data in industry which is cost-consuming.There exists many over-the-top data with longitude/latitude in telco network,these data can be used to label MR data.But,its volume is much smaller than MR data which will cause mismatch problem seriously.We design a map-matching and interpolation method to solve this problem and make labeled data raise from 4 million to100 million.Empirical result shows that localization median error reduce from 95.7m to 83.5m and accuracy increased by 12.7% after enlarging the size of labeled data.(2)Trajectories have strong contextual property,in this paper we improve traditional random forest by proposing a context-aware coarse-to-fine regression model(CCR).CCR is a two-layer cascade random forest.First,We train first layer using coarse raw features and generate fine contextual features according to the outputs of the first layer.Then,combining coarse and fine features to fit second layer.CCR outperforms other state-of-the-art localization algorithms with median error of 83.5m.The accuracy is increased by 12% and 14% when comparing to random forest and standard finger-printing method respectively.(3)In this paper we propse a Hybrid Paritcle Filter(HPF)algorithm.First,HPF will fit a random forest localization model.The state emission probability of HPF is defined as the distribuiton of samples with same label locating in different leaves.Next,state transition probibilty of HPF is calculated from third-party real-world trajectories.Empirical result shows that HPF achieves 75.325 m for median error which outperforms CCR by 15%.Meanwhile,we visualize the prediction of HPF and CCR on map.It shows that the trajectories output by HPF are much smoother than that by CCR,this can prove that HPF has better usage of contextual information than CCR.In HPF,every predicted point is restricted by its previous point,so it won't jump far away.However,although CCR include contextual features in second layer,but random forest only can treat each point independently and unsmooth prediction is inevitable.