Road Closure Detection Based Upon Multi-feature Fusion

With the widespread use of GPS equipment,residents’ daily travel is more and more dependent on map navigation software,in order to ensure the high accuracy of electronic maps,navigation software urgently needs to perceive the dynamic changes of road net-works to update electronic maps.In recent years,a large number of literatures are devoted to the study of the changes of road topology such as missing roads in road networks,de-tection of road intersection location and coverage.However,in the road network,due to traffic accidents,traffic control,road repair and other factors will lead to some of the oth-erwise passable sections of the short-term or long-term can not be normal passage,if not timely detection of such incidents,not only will bring inconvenience to residents travel,but also cause huge economic losses.With the rapid development of location-based ser-vice technology,various data platforms have accumulated a huge amount of trajectory data,as well as rich real-time information on road segments.Practitioners and researchers can use this vast amount of spatial-temporal data to detect road closures in real time and update electronic maps.Traditional navigation software mainly through the service users to provide the function module of road closure real-time reporting to achieve closed road detection,such methods are time-consuming and laborious,and the update cycle is long,can not guarantee the accuracy and timeliness of detection.The existing road closure de-tection research is based on the single characteristics of road traffic flow,which can only identify fully closed roads,can not identify some more complex types of closure,and easy to misjudge traffic congestion as a closed event.In order to solve the above problems,this paper presents a multi-feature fusion road closure detection framework for the first time,consisting of offline closed feature model-ing part and online road closure detection part.In stage of off-line closed features model-ing,this paper first divides the road network equally into grids,and extracts the correlation strength between all roads,then extracts the closed features of the grid and road closure features from the historical data,such as grid’s traffic flow,road’s turning volume and U-turn frequency.The online road closure detection stage is divided into candidate closed grid screening and road closure detection two steps.In view of the spatial-temporal de-pendence of closed features,for different detection objects,combined with convolution neural network,long short-term memory neural network and graph convolution neural network,predict the current time period of each grid(or road)closed feature value,by calculating the difference between each grid(or road)closed feature prediction and the true value,compare them with predefined thresholds,finally determine the closed road.Finally,a large number of experimental results based on three real data in Chengdu,Shanghai and Beijing verify the effectiveness and efficiency of the method.The main work of this article is the following:· Off-line Closed Feature Modeling At this stage,this article first divides the road network into grids equally,and establishes the reverse index of trajectory data and grid numbers.Then,the map matching algorithm based on hidden Markov is used to obtain the road segment sequence corresponding to each trajectory through his-torical trajectory data.Then,the correlation strength between each pair of road segments is obtained based on the statistics of the road segment sequence.Finally,each grid’s traffic flow sequence,road’s turning volume sequence are obtained,and the U-turn frequency sequence on each road is extracted based on the turning point DBSCAN clustering algorithm.· Online Road Closure Detection At this stage,this article first selects the candi-date closed grids based on the grid’s traffic flow through the prediction model based on the convolution neural network and the long short-term memory neural network.Then through the prediction model based on the graph convolutional neural net-work and the long short-term memory neural network,the closed road is detected according to the turning volume of the road and the U-turn frequency.· Road Closure Types Identification At this stage,in order to determine the differ-ent types of closures,this article distinguishes the types of closures based on the periodicity and continuity of closures in terms of time.In terms of space,based on the offset of the trajectory on the closed road,combined with the Wasserstein distance,the special merging road closure type is further distinguished.In summary,this paper presents a multi-featured fusion road closure detection frame-work.The effectiveness and efficiency experiments were carried out on three real trajec-tory data sets,which prove the superiority of this method.