Vehicle Destination Prediction By Considering Historical Travel Patterns And Current Driving Status

Real-time destination prediction is crucial to navigation path planning and location services,and achieving accurate and efficient destination prediction can help users plan the best travel route in real time,optimize travel cost and experience,avoid congested roads and reduce traffic dispatching pressure,etc.Although a large number of scholars have made attempts in the field of destination prediction research,they are still unable to meet the market demand for real-time prediction.On the one hand,existing models lack semantic mining of real-time driving status(e.g.,driving speed,steering,and other real-time operational behaviors),resulting in the inability of models to model external environmental information that affects driving behavior,such as the current road section and road congestion.On the other hand,for real-time prediction models,the influence of the historical travel patterns of mobile objects on the current travel is less considered,which leads to the inability of the models to model the current travel behavior patterns of objects from a global perspective.In fact,the short-term travel behavior and longterm travel pattern of mobile objects are the expression of individual spatio-temporal activity behavior in different time scales,and combining the characteristics of both can better predict travel intention.To address the above two shortcomings,this paper uses the text matching alignment method to mine the key sequence features that are highly correlated with the current travel under the long period of history,and combines the real-time driving features to identify the key trajectory points of the current travel,so as to achieve the effective prediction of the current travel destination of individuals.The main research contents of the paper are summarized as follows:1)Key historical OD sequence characterization based on travel activity chain modeling: In order to comprehensively model the current travel intention of mobile objects,it is necessary to understand their historical travel behavior patterns.In this paper,we construct a travel activity chain based on historical activity events in order to explore the historical travel behavior patterns of mobile objects,with a view to portraying their historical travel characteristics in a long period from the semantic level;at the same time,we identify the sequence segments in historical data that are highly correlated with recent travel behavior based on semantic alignment methods,and then effectively model the current travel intention of the objects.2)Travel driving state based on current trajectory key point identification: realtime driving state records a series of operational feedback of the mobile object on the travel path,which can reflect the surrounding road network structure and traffic congestion condition of the vehicle currently located from the side,and assist in the construction of the current travel pattern.Based on the current travel trajectory data,this paper extracts the driving state features and automatically identifies the key trajectory point locations of the traveled trajectory based on heuristic calculation method,detects and expresses the key spatio-temporal features that are highly relevant to the destination during the user’s travel,and then assists in destination prediction.3)Integration of historical and current travel features: Since the historical travel patterns and current travel patterns of mobile objects differ in terms of time scale and semantic dimensions,they cannot be directly coupled.In this paper,the attention mechanism is used to map the two travel patterns onto their respective spatio-temporal feature vectors to achieve the fusion of historical and current travel patterns,and the predicted destination latitude and longitude coordinates are output using the residual network.This paper conducts experiments based on the trajectory dataset of Shenzhen private car users throughout 2019.The study conducts model validation experiments with four benchmark models: random forest model(RF),long short-term memory neural network(LSTM),attention-aware LSTM for real-time driving destination prediction(LSI-LSTM)considering location semantics and location importance of trajectory points,and individual driving destination prediction(ITP-CMM)considering intersection transfer preferences and current movement patterns The analysis verified the advantages of the models in terms of prediction accuracy and stability.Module ablation experiments were also conducted,and the effects of model internal parameter selection and frequent and infrequent locations on model prediction accuracy were explored.Based on the historical travel activity chain and current travel status,this paper uses attention mechanism,semantic matching alignment and other related deep learning technology methods to achieve effective real-time destination prediction.This research tries to apply the text semantic alignment method of natural language processing to mine and characterize historical activity behavior features in trajectory prediction,and verifies the rationality of applying this method,which provides new ideas and references for related research on mobile behavior feature extraction and characterization,real-time destination prediction,etc.