Objective With the rapid development of visual sensing and artificial intelligence technologies, computer vision is being increasingly applied in fields such as autonomous driving and robot navigation. Visual relative pose estimation can be broadly categorized into monocular and multi-camera systems based on the number of cameras employed. Compared to monocular configurations, multi-camera systems offer distinct advantages, including a wider field of view and richer environmental information through synchronized multi-perspective observation. These attributes render them particularly suitable for applications demanding high safety and robustness, such as autonomous driving, where they have become the predominant sensing configuration. However, existing relative pose estimation methods based on multi-camera setups typically rely on a large number of feature correspondences, resulting in high computational complexity and difficulty in meeting real-time requirements. Especially in the dominant scenario of planar vehicle motion, general methods fail to fully utilize motion constraints, leading to model redundancy and computational resource waste. This study aims to design an efficient minimal solver for multi-camera pose estimation that incorporates planar motion prior, specifically targeting autonomous driving applications on structured roads.

Methods An efficient minimal solver that incorporates prior knowledge of planar motion is proposed. Built upon the generalized camera model, this method fixes the world coordinate system to a spatial point and introduces a depth-parameterized translation representation. In typical autonomous driving scenarios, vehicles predominantly operate on structured roads or near-planar surfaces. This motion characteristic provides a critical constraint for pose estimation: the six-degree-of-freedom rigid motion of the platform can be simplified to a planar motion model involving only yaw rotation about the vertical axis and translation constrained to the horizontal plane. Based on this observation, a planar motion model tailored for multi-camera systems is constructed. This formulation transforms the complex multi-camera pose estimation problem into a linear model that can be solved with only three sets of feature correspondences, significantly reducing reliance on feature quantity and computational complexity. The method is evaluated through comprehensive simulations and real-world experiments, covering performance under image noise, road slope perturbations, and vehicle vibrations.

Results and Discussions The proposed method, named 3pt-Depth, is systematically compared with several baselines including 17pt-Li, 6pt-Stew, and 2pt-Choi in terms of efficiency, stability, and robustness. In computational efficiency, 3pt-Depth achieves an average runtime of 14.755 μs, which is over 3 times faster than 17pt-Li and significantly faster than 6pt-Stew (Tab.1). Under noise-free conditions, 3pt-Depth exhibits competitive numerical stability, with translation error performance better than 17pt-Li and clearly superior to 6pt-Stew (Fig.3). Simulations are conducted using a binocular vision system with a baseline length of 1 meter by default, consistent with sensor configurations in the widely used KITTI dataset and representative of real-world automotive setups. To systematically evaluate method performance under various disturbances, we examine the impact of image noise, road slope, vehicle vibration, and calibration errors. All solvers are integrated within the RANSAC framework, and the estimate yielding the highest inlier count is used for final error evaluation. In image noise tests, 3pt-Depth demonstrates the strongest tolerance to image noise across different motion patterns, with the slowest accuracy degradation as noise increases (Fig.4). It also outperforms 2pt-Choi under road slope disturbances (pitch and roll) and vehicle vibration scenarios (Fig.5-7). Further experiments show that 3pt-Depth maintains accuracy under increasing extrinsic rotation and translation calibration errors, and scales well with the number of cameras (Fig.8). Real-world validation on the KITTI dataset shows that 3pt-Depth achieves lowest rotation and translation errors across all sequences, confirming its practical effectiveness in autonomous driving environments (Tab.2-3).

Conclusions This paper presents an efficient minimal solver for multi-camera pose estimation under planar motion constraints, specifically designed for autonomous driving. By leveraging the planar motion prior and a depth-parameterized translation representation, the method requires only three feature correspondences, significantly reducing computational complexity and dependency on feature matches while maintaining estimation accuracy. The method shows strong robustness against common disturbances in driving environments. Future work will focus on extending the method to adapt to non-ideal planar motions and exploring tight-coupling frameworks with other sensors like IMUs to enhance robustness in complex real-world scenarios.