Monocular vehicle pose estimation based on 3D model

Vehicle pose estimation is an important component of intelligent transportation systems. However, the complex scenes and loss of depth information are challenging problems in the estimation. This paper proposes a method that combines monocular pose estimation and a 3D vehicle model to estimate vehicle pose. First, a multi-scale vehicle are normalized, and then the coordinates of key points are predicted in the form of a vector field to increase the accuracy of the pose estimation for the truncated and occluded vehicle. Furthermore, a distance-based loss function for the vector field and key point error minimization voting method is established to further improve the accuracy of the pose estimation algorithm. In addition, we propose a synthetic vehicle pose estimation dataset with rich annotation information. The verification results show that the average position and angle errors of our algorithm are 0.162 m and 4.692°, respectively. Our method provides significant improvements over existing methods and has considerable practical application value. © 2021 Science Press. All right reserved.