Reproducing traffic loads on long-span bridges via fusion of sparse cameras and weigh-in-motion information

Accurate reproduction of the spatial-temporal distribution of traffic loads on bridge decks is significant for bridge health assessment. Current traffic load monitoring systems based on computer vision have the problems of limited field of view coverage and poor accuracy in target matching across multiple cameras. Therefore, the application of these systems to the entire deck area of long-span bridges still faces challenges. This paper proposes a method that combines sparse camera information with vehicle trajectory prediction to reproduce the spatial-temporal distribution of traffic loads. First, the vehicle weights recorded by the weigh-in-motion (WIM) system and the vehicle trajectories identified from the videos are integrated to generate the vehicle load distribution within the range of cameras. Next, an Informer-based deep learning model is utilized to identify driving intentions and predict the vehicle trajectories in the blind areas of the cameras. The efficiency and reliability of the proposed method are verified through the cable-stayed bridge field test. The results show that the identification accuracy of vehicle types is about 91.5%. Additionally, the root mean square errors (RMSE) of the vehicle’s longitudinal and transverse positions are less than 0.9917 m and 0.2100 m, respectively. The proposed method saves monitoring costs while enhancing the robustness of the monitoring system. Furthermore, this method expands the application scenarios to long-span bridges and provides an accurate spatial-temporal distribution of traffic loads for the subsequent real-time bridge health assessment.