Sparse Spatio-Temporal Trajectory Modeling For Monitoring And Prediction

The spatio-temporal trajectory data is essentially information containing time-varying geographic positions.The increasing popularity of location-gathering devices makes more and more trajectory data available,which brings opportunity to intelligent transportation and other fields.It enables us to make real-time monitoring and prediction of road conditions.Thus we can alleviate traffic congestion in big cities.However,some inherent characteristics of this data make traditional methods ineffective in practical applications.To solve unlabeled data problem and data missing in trajectories,this paper has two innovations as follow.1.We propose a probabilistic graphical model for unlabeled trajectory monitoring,which models characteristics in unlabeled trajectory data.It is suitable for data with unknown latent variables and indirectly obtained observations.The learning algorithm based on iterated conditional modes utilizes the relationship between observations and latent variables,amd makes updates alternately.Thus it infers observations and latent variables simultaneously.2.We propose a neural network model for sparse trajectory prediction,which is mainly used to solve sparsity in trajectory data.It is suitable for the datasets which contain a lot of missing values.This model focuses on filling missing values and making predictions at the same time.It updates parameters through co-training algorithm in order to avoid the cascaded propagation of error.These methods are verified in intelligent transportation field.The probabilistic graphical model is applied to travel time allocation.We model the correlation between traffic lights in a fine-grained manner to get accurate time consuming.The neural network model is applied to road speed prediction.We use matrix factorization to update the missing values for accurate future speed prediction.The experimental results show that these methods perform very well in real applications.They overcome some inherent defects of spatio-temporal trajectory data and make up for the shortcomings of existing methods.