Trajectory Data Repair Based On Bidirectional Recurrent Neural Network

In the wave of smart city construction,there are many applications related to monitoring and location-based services,such as traffic flow monitoring,next location prediction,hotline recommendation and so on.These applications need to obtain the geographic information of the moving entities and track their trajectories.With the rapid development of technology,the popularity of devices such as video surveillance systems,GPS,and smart phones has made it easy for us to acquire massive trajectory data of the moving entities.However,under natural conditions,the interference of environmental factors reduces the quality of the original trajectory data.What's worse,the low quality data will reduce the effect of above practical applications.Therefore,it is necessary to adopt reasonable technical to repair the trajectory data and improve the trajectory data quality.For the problem of trajectory data repair,the existing research results mainly include three technologies:map matching,rules based on road network and methods based on historical data.One of the characteristics of trajectory data is its temporal-spatial sequence.However,among the current research results,map matching and road network based methods only use the spatial sequence of trajectory data and ignore its time sequence property.For this reason,this paper proposes a reasonable and effective method to solve the problems of illegal track record and missing track record in the trajectory data caused by the failure of the moving entity ID identification and track record acquisition.First of all,we present an effective data completion method,which consists of two steps.1)We make full use of the noise data generated by the failure recognition of the moving entity ID,and filling some missing values based on the moving entity ID edit distance,time limit conditions,and speed limit conditions of the moving entity.2)Finding similar trajectories in the dataset and using the complementary traits of similar trajectories to further filling partially missing values.Then,we use word2vec to vectorize the location to better explore and use the inner relationship between the locations.Finally,we select the optimized BRNN model that has achieved great success in processing time series data,and the location vector is used as an input to the BRNN model to complete the final trajectory repair.We implemented the word2vec and BRNN model on TensorFlow and conducted extensive experiments using real traffic data sets.The experimental results fully demonstrate the validity and advancement of our method.