Trinocular camera self-calibration based on spatio-temporal multi-layer optimization

In stereo vision systems with dynamically rotating cameras, the accuracy of camera self-calibration method is reduced due to the interference of space noise and mismatched features. To address this issue, a new self-calibration method for trinocular camera is proposed. Firstly, to obtain the uniformly distributed and high-quality matched feature points required for initial calibration of camera pose, according to the mapping relationship between the three-view feature matching points, a three-view grid feature support estimator is defined, and a three-view ring matching method with double-layer feature closed-loop verification is designed. Then, according to the projection relationship of spatial feature points, a new heterogeneous cross-projection optimization function based on closed-loop features is established, achieving accurate calibration of trinocular camera system. Comparison experiments of multiple scenes verify the effectiveness of the method, particularly in the low-textured scenes, where the average Sampson error ranged between 1.37e-11 and 0.061 pixel. Furthermore, the proposed method achieves higher calibration accuracy than the comparative method, which can improve the robustness of dynamic rotating cameras under spatial noise conditions.