| description abstract | Improving the efficiency and safety of highway traffic relies heavily on accurately modeling the complex dynamics of traffic flow. This study aims to develop a novel method for modeling highway traffic flows by integrating data from multiple sources, such as roadside cameras, gantry cameras, and communication devices. The method leverages heterogeneous sensor data characteristics, temporal information, and spatial structures to achieve deep fusion of latent features at the sensor level. To minimize human intervention, a multistage training approach is employed, combining large-scale self-supervised learning with supervised fine-tuning, leveraging abundant unlabeled unstructured data such as monitoring videos recorded by roadside cameras and snapshots captured by gantry cameras, alongside limited accurate structured traffic flow data aggregated by communication devices from gantries and toll stations. We demonstrate the effectiveness and stability of the proposed method on a case study, the G92 ring highway around the Hangzhou Bay in Zhejiang Province, China, achieving the mean absolute percentage error of traffic flow within 7.5% and 8.3% for fixed and variable highway sensor networks, respectively. Ablation studies further demonstrate the significant improvement in predictive accuracy achieved by the designed self-supervised pretraining task. To summarize, our approach provides a promising solution for efficient and safe management of highway traffic flow, with potential applicability to real-world scenarios. | |
