Expressway traffic state recognition based on multi-source data fusion and multi-view fusion graph neural network under velocity feature mapping

To comprehensively extract the time series features of average vehicle velocity data on expressways and their correlation with traffic states, this paper proposes a Multi-view Fusion Chebyshev Graph Convolution Network (MvFCGCN) model for accurately recognizing expressway traffic congestion states. Firstly, we propose a weighted fusion method of checkpoint data and radar velocity data to obtain the traffic state feature vectors, mapping them into heat maps in the form of chromatograms to create the Traffic State Feature Image dataset based on Checkpoint-Radar Data Fusion (TSFI-CRDF dataset). Secondly, a Traffic State Deep Clustering Network (TSDCN) model based on multi-view fusion convolutional neural network and variational autoencoder is constructed to automatically classify and label the traffic state feature images in the TSFI-CRDF dataset. Subsequently, the traffic state feature image data is further mapped into graph structure data, and the MvFCGCN model is constructed based on the Chebyshev graph convolutional neural network with integrated view fusion weights for traffic state recognition. Finally, experimental validation is carried out on the example of checkpoint plate recognition data and radar velocity data collected from the Beijing-Qinhuangdao section of the Beijing-Harbin Expressway. Comparative analyses with models such as Convolution and Self-Attention Network (CoAtNet) are performed, as well as ablation experiments, alongside effect analyses of the TSFI-CRDF dataset. The experimental results demonstrate that the MvFCGCN model achieves an overall recognition accuracy of 95.25 %, outperforming other comparison models. The proposed interpolation method for fusion of checkpoints and radar data effectively restores the original velocity feature of the traffic state.