Trajectory Prediction Of Moving Objects With Multi-feature Fusion Based On Deep Learning

With the massive increase in global navigation satellite system(GNSS)trajectory data of moving objects and the rapid development of urban road traffic systems,road traffic monitoring and prediction issues play a vital role in urban traffic management.Among them,the accurate trajectory prediction of moving objects has become an important research direction,which is the key technology affecting the application of urban traffic flow monitoring and unmanned vehicle scheduling control.However,most of the existing prediction methods only consider the single temporal feature or spatial feature of the trajectory data,so the prediction accuracy is limited.In this paper,trajectory prediction of moving objects based on GNSS data is deeply studied.Theoretical research and application are carried out from statistics-based methods and neural network methods.The multi-feature trajectory prediction method that incorporates the temporal and spatial characteristics of the trajectory can use the two-dimensional spatial information of the traj ectory and extract the temporal characteristics of the trajectory.At the same time,with the assistance of additional multiple features,the feature richness of the trajectory is retained to the greatest extent,which improves the accuracy of trajectory prediction.The main work and contributions of this paper are described as follows:1.Aiming at the low efficiency of the existing GNSS trajectory feature extraction methods,this paper proposes a trajectory spatio-temporal feature extraction network based on deep learning.The network includes two parts:spatial feature extraction and temporal feature extraction.For the extraction of spatial features,a generation algorithm is proposed which can transform one-dimensional trajectory sequences to two-dimensional trajectory images.Then a convolutional neural network(CNN)with two convolution layers and two pooling layers is used to extract the spatial features of different depths of the trajectory image.For the extraction of temporal features,the original one-dimensional trajectory is cut into sub-trajectories at first.Then a bidirectional long and short-term memory network(Bi-LSTM)with single hidden layer is used to extract the temporal characteristics of the trajectory.Finally,the spatial and temporal features are spliced into spatio-temporal feature vectors.Experiments were conducted on the real taxi data set of Porto.The results show that the proposed image generation algorithm can effectively preserve the spatio-temporal information of the original trajectory.Compared with the image generation algorithm in T-CONV(Trajectory Convolution),the image generation rate is basically the same.The integrity of the trajectory points in the trajectory map is improved by 111.9%on average,and the density of the trajectory image is also improved by 61.3%.2.Aiming at the problem that the existing trajectory prediction methods are difficult to fuse multi-source heterogeneous features,this paper proposes a trajectory prediction model based on heterogeneous feature fusion of text and GNSS data.First,a word embedding algorithm is used to construct different feature vectors from the trajectory data.A trajectory portrait generation model is used to construct trajectory classification vectors.In addition,for special text data,more complex auxiliary features are obtained through the name entity recognition(NER)model.Through the multi-feature fusion model proposed in the paper,vectors of different dimensions are integrated,which increases the diversity of prediction features.Experiments were conducted on the real taxi data set of Porto.Compared with methods based on statistics,the prediction error is reduced by more than 10%.Compared with methods based on neural network such as MLP,RNN and T-CONV,the error is reduced by 4.9%on average.3.Combining spatio-temporal feature extraction model and the multi-feature fusion model,this paper constructs a complete trajectory prediction framework TCL(Traj ectory-CNN-LSTM).The model accepts one-dimensional trajectory data as input and outputs specific prediction coordinates.To verify the performance of TCL framework,an integrated presentation platform is built.The system integrates three modules:trajectory prediction,geographic noun extraction of text trajectory and marking visualization,and trajectory portrait visualization.The system realizes the dynamic demonstration of local trajectory data.