Research On Traffic Flow Data Recovery And Prediction Method Based On Road Network Pixelization

With the accelerated process of urban informatization,the number of sensors in the road network increases year by year,and traffic data shows explosive growth.However,due to the limitation of construction cost and construction time,the quality of sensors in the road network varies,and the problem of data loss inevitably exists in data collection,transmission,and storage.High-quality traffic data is the basis for constructing the Intelligent Transportation System(ITS).The missing data brings great challenges for the subsequent work,which is a problem that needs to be solved for the construction of intelligent transportation systems and the development of smart cities today.Traffic flow data prediction is an important application of the intelligent transportation systems.Realtime and accurate traffic flow data prediction is significant in reducing traffic jams,improving traffic efficiency,and improving resource utilization.Complete and reliable traffic flow data can better express the spatio-temporal and periodic features of complex road network data,which is an important basis for improving the accuracy of traffic flow data prediction.This paper focuses on traffic flow data representation,traffic flow data recovery,and traffic flow data prediction.The preliminary work and contributions of this paper are as follows.1.In the aspect of traffic flow data representation,a representation learning-based method for traffic flow data representation is proposed to convert raw traffic data into pixelated representation,which solves the heterogeneity of traffic data and video data.First,a road network pixelation algorithm is proposed to mine the spatial-temporal and traffic correlation of road network checkpoints from the massive traffic trajectories and map the road network checkpoints into the road network pixel matrix.The corresponding checkpoint traffic is filled into the pixelated matrix to generate road traffic flow images by slicing the traffic data in the road network.Secondly,to generate a better road network pixel matrix,this paper focuses on the purpose of road network pixelation and proposes a road network pixel matrix evaluation method to evaluate the road network matrix generated by unsupervised methods.2.In traffic flow data recovery,an end-to-end traffic flow data recovery model based on road network pixelation is proposed and designed to extract spatio-temporal features from sparse road traffic flow images with multiple missing patterns.First,to extract spatio-temporal features from the sparse pixelated images of the road network.In this paper,Partial Convolutions and LSTM are applied to the extraction of traffic flow features,and a network unit PConv LSTM applicable to traffic flow recovery is proposed for extracting the periodic features of traffic flow.Secondly,this paper tries to apply the video inpainting technique to the traffic flow data recovery,and uses two sub-networks to extract the periodic and spatio-temporal features of traffic flow for traffic flow data recovery,respectively.3.For traffic flow data prediction,a traffic flow data prediction model based on pixelation of road networks is proposed and designed to realize spatio-temporal feature extraction of traffic data near moment and period data.It also fuses the influence of external factors such as holidays to implement the traffic flow data prediction.First,Partial Convolutions GRU is used to extract traffic flow features for the complex spatiotemporal and periodic features of traffic flow.Second,considering that various external factors influence traffic flow,an external data fusion network is proposed to fuse holiday and weather information,which has a large impact.Finally,an end-to-end traffic flow data prediction model is designed for the characteristics of traffic flow data prediction tasks to achieve real-time and accurate traffic flow data prediction,and to complete the closed-loop of traffic flow data from representation to recovery and finally to prediction.This paper conducts relevant experiments using ANPR data collected in real environments for the above primary work.The results show that the proposed method achieves higher accuracy in traffic flow data recovery and prediction than the baseline method.