Research On Crowd Stationary And Traffic Spatio-temporal Pattern At City Scale

With cities becoming large-scale,higher requirements have been put forward for city managers.City planning,transportation,energy saving,environment protection all depend on real-time and reliable data for scientific decision-making.Based on the data retrieved from mobile networks,intelligent transportation systems and check-in data from social networks,this paper focus on statistical analyzing and modeling behaviors of human mobility and functional regions.We use statistics,machine learning,pattern recognition and natural language processing methods to discover city’s spatio-temporal patterns.It consists of four parts.Firstly,the spatio-temporal pattern of urban residents and regions are analyzed from macro-scale.The gathering patter and mobility pattern are studied with spatial point pattern analysis and time series analysis.Hotspots distribution,human flow change rate and tidal effect are studied with grid method.Afterwards,the concept of regional difference index is proposed and the spatio-temporal correlation between regions is studied.Consequently,two kinds of spatio-temporal patterns are analyzed in detail,including crowd stationary pattern and traffic pattern.For crowd stationary pattern,the stationary trajectory segments are extracted firstly.Since the recorded trajectory points are sparse,the entering and leaving time of each segments are estimated using the weighted least square method with the combination of user-specific pattern,the region-specific pattern and the global pattern,which is proved to be accurate.Afterwards,a hierarchical Bayes model for clustering human behavior is built,which is solved by the joint usage of expectation maximization algorithm and Markov chain Monte Carlo algorithm.It can effectively discover crowd stationary pattern and advantageous in model evaluation metrics.For traffic pattern,the spatial and temporal patterns of metro and taxi traffic are discovered and compared with the method of tensor decomposition.Leveraging the distribution of POI,an algorithm based on deep neural network is proposed for traffic tensor reconstruction,which can be used for newly built regions without any history traffic information or regions to be reformed,and is proved to be relatively accurate.Finally,a novel topic model is proposed,which can distinguish between global and local topics,effectively filter out noises and discover spatial topics and hence are useful to understand the patterns discovered.