| As an important urban infrastructure,urban road-network is closely related to people's daily life.With the continuous growth of urban population and car ownership,urban road-network is bearing increasingly severe load pressure,resulting in frequent traffic congestion and traffic accidents,which seriously restricts the development of the city.Therefore,real-time perception and prediction of urban road-network status and identification of high-risk areas of urban geospatial traffic accidents are of great significance to improve urban road-network operation efficiency and traffic safety.With the development of sensor,internet of things,crowd-sensing,a large amount of urban multi-source big data will be generated in the process of urban operation: vehicle trajectory data such as taxis and buses;induction coil,checkpoint,surveillance camera and other road monitoring data;temperature,air quality,weather and other meteorological data;mobile signaling data and check-in data of various applications.These multi-source data are interrelated.Combined with big data and artificial intelligence related algorithms,they can extract the potential operation mode of urban traffic,so as to provide data and theoretical support for the construction of urban intelligent transport system(ITS),and provide data and theoretical support for improving transportation efficiency,alleviating traffic congestion and reducing traffic accidents improve urban planning and provide solutions.However,the existing researches in the field of intelligent transportation system still face the following challenges:(1)Due to urban big data has different sources,it is difficult to obtain full data from different sources and different types in the same time period,resulting in certain one sidedness and limitations in the research results;(2)Massive data requires high data processing capacity,consumes a lot of storage space and computing resources,and is difficult to meet the real-time requirements,especially for short-term prediction tasks;(3)In addition,the potential association between multi-source data is difficult to capture,and there are still some challenges in how to build an effective data fusion scheme.Based on this,this paper aims to overcome the above challenges and build basic models of ITS.Finally,it puts forward the urban road network condition monitoring model,the urban road network traffic flow prediction model and the urban geospatial traffic accident risk inference model.Specifically,the main research contents of this paper are as follows:(1)In this paper,an urban road network condition monitoring model based on the fusion of mobile phone signaling data and taxi trajectory data is proposed to solve the problems of limited coverage and data deviation in the current road network monitoring system researches.Firstly,according to the characteristics of "ping-pong switching" and "data drift" in mobile phone signaling data,a preprocessing method is proposed to filter,modify and screen the original signaling data.On this basis,the historical GPS tracks of 7000 taxis are fused to interpolate the mobile phone signaling data in time and space,so as to solve the problem of data sparsity.Next,a road network matching algorithm based on Hidden Markov Model(HMM)is proposed to convert the observed mobile phone signaling sequence into the user's real trajectory in the road-network.Finally,the average speed of all roads in urban road-network is calculated by the speed estimation module,and realize the real-time monitoring and visualization of urban road network traffic conditions.(2)In this paper,a traffic flow prediction model based on historical traffic monitoring data of a limited number of roads is proposed,which reduces the demand for storage space and computing resources.By selecting a limited number of monitoring roads and taking their historical traffic data as input,the traffic flow prediction of all roads in the urban road network is realized at a low cost.Firstly,according to the spatial structure information and attribute information of each road in the road-network,each road is transformed into an embedding vector.On this basis,all roads are clustered by K-means algorithm,and the central node of each cluster is used as the initial set of monitoring roads.Then,through the monitoring road selection algorithm proposed in this paper,L monitoring roads are selected iteratively,in which L is far less than the total number of all roads.After that,the historical data of the selected L monitoring roads are used as the input of the multi-head attention mechanism module,and the external characteristics around the target road(including weather and POI distribution)are integrated to obtain the correlation between the target road and each selected monitoring road under different space,time and external factors.Finally,the prediction results are obtained through two full connection layers.(3)By fusing easily available data,including urban satellite images,human mobility data,traffic data,POI distribution data and urban road network,this paper proposes a pixel level fine-grained traffic accident risk inference model based on the fusion of satellite images and multi-source data.Specifically,taking the multi-channel image composed of RGB satellite image and road network binary image of the target area and the target neighborhood(the doubled target area,which is used to extract the edge information of the target area)as the input of the model,the characteristics of different layers are captured through two VGG-16 networks,Then,multi-scale and different visual receptive field features are captured by atrus convolution.On this basis,the characteristics of different levels are fused through the spatial attention module,and the characteristics of external factors such as human mobility,traffic condition and POI distribution in the target area are fused through the channel attention module.Finally,the image with the same size as the target area is obtained,and each pixel represents the evaluation value of traffic accident risk at the corresponding position.The main innovations of this paper are as follows:(1)In the proposed urban road network condition monitoring model,by taking advantage of the large number of users and wide coverage of mobile signaling data,it solves the limitation and one sidedness of urban road-network operation condition monitoring based on single traffic mode data.The availability of mobile signaling data in urban road-network monitoring task is improved by spatio-temporal fusion interpolation with high-precision vehicle GPS trajectory.(2)In the proposed traffic flow prediction model,the roads are transformed into vectors through the spatial structure and attribute information of the road-network,and a limited number of monitored roads are selected,with their historical traffic data as input,the traffic flow prediction of all roads is realized,which reduces the demand for storage space and computing resources,and has higher practicability.(3)Compared with current traffic accident risk prediction models based on urban grid or the whole road section level,the traffic accident risk inference model proposed in this paper realizes pixel level fine-grained traffic accident risk inference by using easily available urban multi-source data,which is helpful to provide more reasonable traffic accident prevention guidance measures.And the model has high practicability and scalability,which is of great significance to improve the universality and practicability of traffic accident risk prediction at a lower cost in practical application.To sum up,this paper integrates urban multi-source big data and proposes a series of intelligent transportation related models using the combination of multi-source data and artificial intelligence algorithms.All models are trained and verified by the real dataset of Changchun City.Through a large number of experimental results,this paper proves the effectiveness of the models,and provides important solutions and technical support for enhancing traffic supervision,alleviating traffic congestion and improving traffic safety. |
PDF Full Text Request