Urban Road Network Flow OD Estimation And Prediction Using Attention-Based Deep Neural Networks

The road network flow OD matrix describes the spatiotemporal distribution of vehicle travel in urban road networks and provids important data support for urban road traffic planning and management.Early methods of obtaining static OD flow based on manual surveys had the problems associated with high data collection costs and untimely data updates,making it difficult to meet the needs of current dynamic traffic management.Existing mainstream dynamic OD estimation methods based on reverse inference of road segment traffic flow typically involve two complex optimization problems:dynamic OD estimation and dynamic traffic assignment,and the modeling and solving process is often extremely challenging.With the increasingly advanced traffic detection technology,the collection of diverse and refined traffic big data has become easier.Under this background,the estimation and prediction of road network flow OD based on refined vehicle travel trajectory data has gradually become one of the mainstream and emerging methods for obtaining dynamic traffic demand data for the road network.Despite this,accurately and reliably estimating and predicting road network flow OD using vehicle trajectory data is not an easy task.On the one hand,due to factors such as traffic detection equipment failures,data transmission or storage errors,the obtained vehicle travel trajectory data is often incomplete in time and space,presenting fragmented vehicle travel trajectory segments.How to accurately reconstruct vehicle travel trajectory has become an extremely critical and challenging issue in the research of road network flow OD estimation based on vehicle trajectory data.Existing road network vehicle trajectory reconstruction methods mainly rely on empirical assumptions about vehicle path selection behavior(e.g.,assuming that vehicles usually travel the shortest travel time path on the road network).When the assumption conditions do not match the actual situation,the accuracy of the vehicle trajectory reconstruction model will significantly decrease,which will further affect the accuracy of OD estimation.On the other hand,considering that the time-varying OD matrix characterizes the dynamic spatiotemporal distribution of road network traffic demand,how to accurately characterize and capture the complex nonlinear spatiotemporal correlation relationship underlying road network flow OD flow data has become a key challenging issue in current road network flow OD prediction modeling.Given the above background,based on the automatic license plate recognition data and the trajectory data extracted from unmanned aerial vehicle(UAV)videos,this paper aims to study the method of estimating and predicting urban road network traffic flow origindestination(OD)using vehicle trajectory data driven by deep neural network algorithms and multi-head self-attention mechanisms.The focus is on two modeling works:the first is to construct an OD estimation model based on a multi-head attention deep neural network(MHA-DNN),and the second is to construct an OD prediction model based on a multi-head attention long short-term memory neural network(MHA-LSTM).The main research work and conclusions of the paper include the following three aspects.(1)For vehicle trajectory data preprocessing:Firstly,for automatic vehicle license plate recognition data,targeted preprocessing strategies are designed to carry out four aspects of preprocessing work,including data deduplication,data validity check,travel time anomaly check,and data mapping matching.Secondly,for vehicle trajectory data extracted from unmanned aerial vehicle aerial video,targeted preprocessing strategies are designed to carry out four aspects of preprocessing work,including data structure conversion,zero-speed drift check,data filtering,and longitude and latitude drift check.Finally,a vehicle travel trajectory chain division method is proposed for the above two types of trajectory data sources,and a vehicle travel trajectory chain dataset is constructed based on this,providing data support for subsequent road network traffic flow OD estimation.(2)For estimating road network flow OD:A urban road network flow OD estimation method based on MHA-DNN is proposed,including three core technological steps.First,based on the complete vehicle travel trajectory chain dataset,a vehicle trajectory reconstruction model is established.By introducing a multi-head self-attention mechanism,the model focuses on representing the interdependent relationships between nodes within the vehicle trajectory sequence,and utilizes the complex non-linear feature extraction and pattern recognition capabilities of deep neural networks to learn and capture the underlying travel laws on the road network from the vehicle travel trajectory chain dataset.Secondly,the established vehicle trajectory reconstruction model is used to complete the missing vehicle travel trajectory chains and obtain the complete travel trajectory chains of all detected vehicles.Finally,based on the detection rate of vehicles at road network intersections,a sample expansion strategy is designed to infer the flow OD.The effectiveness of the proposed vehicle trajectory reconstruction model and the rationality of the road network flow OD estimation method were evaluated and validated using automatic license plate recognition data from Kunshan City,China,and vehicle trajectory data extracted from unmanned aerial vehicle video in Athens,Greece.The results show that by using trajectory data with sampling rates of 50%,70%,and 80%,the proposed vehicle trajectory reconstruction model outperforms typical KNN and Text-CNN models.In addition,through accuracy evaluation of reassigned link flows and visualization analysis of the estimated road network flow OD,the rationality of the proposed OD estimation method was validated.(3)For road network flow OD prediction:A method based on MHA-LSTM is proposed.By introducing multi-head self-attention mechanism,the method fully focuses on representing and encoding the interdependent relationship between time-varying road network traffic flow OD matrices.On this basis,long-short-term memory neural network is utilized to capture the complex nonlinear correlation between the encoded time-varying traffic flow OD matrices,thus achieving reliable prediction of road network flow OD.The performance of the proposed method and the comparative methods were evaluated and verified using road network flow OD data sets obtained at different time intervals in Kunshan,China and Athens,Greece.The results showed that the proposed method outperformed the comparative methods in terms of prediction performance at all selected time intervals.