An accurate approach for obtaining spatiotemporal information of vehicle loads on bridges based on 3D bounding box reconstruction with computer vision

Abstract

Vehicle load is an important basis for bridge design, condition assessment, and maintenance and strengthening, and its statistical laws can help us further understand the behavior of bridges under vehicle loads. Therefore, it is important to obtain accurate vehicle weight and vehicle spatiotemporal information that reflects the load time history. As of now, vehicle weight can be obtained using WIM technology, but there are still some issues in the methods of obtaining vehicle information. The coordinate transformation method is simple to operate, but sacrifices accuracy. The existence of a certain distance from the tail of vehicles to bridge decks makes the results based on dual-target detection differ from the actual. Therefore, an accurate approach for obtaining spatiotemporal information of vehicle loads on bridges based on 3D bounding box reconstruction with computer vision is proposed in this paper. To achieve this, a deep convolutional neural network (DCNN) and the You Only Look Once (YOLO) detector are used to detect vehicles and get the 2D bounding box. By establishing the relationship between 2D and 3D bounding box of the vehicle, an algorithm for 3D bounding box reconstruction of vehicles is proposed to get the sizes and position of vehicles. The spatiotemporal information of the vehicle loads is finally obtained by using multiple objects tracking (MOT). To verify the accuracy and reliability of the proposed approach, a bridge vehicle loads identification system (BVLIS) was developed and tested on a cable-stayed bridge in operation. The results show that the approach is accurate and reliable, and can be used to obtain vehicle information and provide vehicle load boundary conditions for bridge finite element modeling.