The existing pose measurement methods in monocular visual guidance are mainly divided into cooperative methods and non-cooperative methods. In the cooperative method, the pose parameters between platforms are measured by laying cooperative signs on the target platform. In the non-cooperative method, the target features and image information are used to calculate the pose parameters between platforms. Both cooperative and non-cooperative methods need accurate 3D models. And, the high-precision pose measurement cannot be achieved with the absence of accurate 3D model. Therefore, drawing on SLAM(simultaneous localization and mapping) and SFM(structure from motion), this paper proposes to optimize pose and 3D model iteratively by using multiple view geometry constraint information in sequential images, so as to achieve high-precision pose measurement and reduce model errors.

In this paper, the target’s 3D model is expressed as a set of sparse 3D keypoints. And the 2D projections of 3D keypoints in a single image are detected by RTMPose(real-time multi-person pose estimation) algorithm, which is commonly used in the field of human pose estimation. Further, the initial 6D pose is calculated by solving the PnP(perspective-n-point) problem combined with the inaccurate 3D model. Then, the 3D keypoint position is expressed in the form of linear parameter equation. The optimization objective function is established based on the object-space collinearity error, and the target’s 3D model and 6D pose are iteratively optimized. By solving the optimization problem, the high-precision pose can be solved with and the target’s 3D model can be reduced effectively.

In this paper, the performance of the proposed method is explored under the background of the rendezvous mission of aircraft landing. Specifically, this paper uses blenderProc to simulate the rendezvous scene of aircraft and ships, and corresponding images are tendered simultaneously. In experiments based on simulation data, the proposed method is compared with RTMPose-EPnP() and PP-TinyPose(paddle paddle tinypose)-EPnP(efficient perspective-n-point) methods. Experiment results show that the proposed method achieves higher precision pose measurement results in the entire image sequence compared with the two comparison methods. At the same time, benefiting from iterative optimization, the proposed method can also effectively reduce the error of the target’s 3D model.

Aiming at the problem of pose measurement between target platforms with the absence of accurate 3D model in monocular vision guidance, this paper iteratively optimizes target’s 3D model and pose. And a new monocular vision measurement method is proposed. Specifically, a set of sparse 3D keypoints is adopted to represent the target’s 3D model, and an advanced keypoint detection method is adopted to achieve robust and efficient keypoint detection. Further, using linear parametric equation form to represent 3D keypoints, and the target’s 3D keypoints and 6D pose are taken as parameters to be optimized. The optimization objective function is established based on the object-space collinearity error. By solving this optimization problem, the target’s 3D model and pose are iteratively optimized, and the high-precision pose is solved. To meet the real-time and online application requirements of pose measurement in visual guidance, sliding window constraints were adopted to limit the scale of calculation. Additionally, the keyframes were selected to reduce the amount of computation and information redundancy. The experimental results show that this research can achieve real-time and online high-precision pose measurement, and effectively optimize the target’s 3D model.

Although the proposed method has achieved obvious pose and 3D model iterative optimization effects, it still has the problem of not adapting to large 2D keypoint detection errors. In the next step, we will explore the use of multiple view geometry constraint information of sequence images to further reduce 2D keypoint detection errors.