Vehicle Category Prediction Based On Historical Trajectory

With the dramatic increase in people's travel needs and the rapid growth in the number of various types of vehicles,the analysis and mining of vehicle trajectory data has received extensive attention.Thanks to the popularization and application of intelligent transportation systems in major cities,a large amount of vehicle trajectory data can be recorded and stored.By analyzing and mining the historical trajectories of these vehicles,many important applications can be generated,such as predicting the next location of some vehicle,discovering the movement pattern of the specific type of vehicle,and the like.In these applications,vehicle category prediction attempts to predict the category to which a vehicle belongs based on the historical trajectories of the vehicle.Adding reasonable vehicle category annotations to a wide variety of complex trajectories,enabling other trajectory mining applications to obtain a more comprehensive understanding of these trajectory data and thus achieve better performance.To study the problem of vehicle category prediction based on historical trajectory,the key is to extract the features of vehicles from their historical trajectories.Existing feature extraction methods mainly include artificial feature extraction,matrix decomposition,or tensor decomposition.Early research methods focused on artificial feature extraction,which required knowledge of specific fields and relied on experts in specific areas to manually extract and select features from the trajectories of vehicles.After that,by counting the frequency of vehicles visiting each location from their historical trajectories,a matrix or tensor reflecting the preferences of different vehicles for each location is obtained.Drawing on the idea of collaborative filtering,the features of vehicles are obtained by decomposing the matrix or tensor.The features of the vehicle obtained in this way only extract features from the visiting preference matrix,ignoring the temporal-sequential correlation information contained in the original trajectory data,and thus its performance on the specific vehicle category prediction task is unsatisfactory.With the extensive application of deep learning in the computer field,some unsupervised feature learning models based on neural networks have begun to attract the attention of researchers.Thanks to the ability of deep learning methods to extract high-dimensional features,these models can learn more comprehensive and rich features contained in the original data.Based on these related studies,we propose a double-view feature learning model for vehicles.The model constructs a feature learning model from the perspectives of both vehicle and location,and learns the features of vehicles in an unsupervised manner.First,from the perspective of the vehicle,based on the historical trajectory of a given vehicle,we build a model to predict which location a vehicle has visited or will visit in a certain time frame before and after a certain location.Secondly,from a location point of view,we build a model based on a visiting vehicle list of some location,building a model to predict which vehicle will visit this location within a certain time before and after a certain vehicle.Then,by combining two feature learning models and sharing the features of vehicles and locations,a final double-view feature learning model for vehicles is formed,and this model is trained using alternate training strategy and stochastic gradient descent method to obtain the features of vehicles.Finally,the results of vehicle category prediction can be obtained by applying these feature vectors to the classifier.We have applied the proposed model and existing benchmark models together on two real datasets and conducted a large number of comparative experiments.The experimental results show that the double-view vehicle feature learning model has achieved excellent performance in the vehicle category prediction task and is superior to the existing models on the given evaluation metrics.Thus we have confirmed its correctness and effectiveness and we think that the model has certain application value and can be promoted and applied in practice.