Abstract:

The existing visual SLAM frameworks based on nonlinear optimization， in order to reduce the computational resource occupation and improve the system operation speed when solving large-scale linear equations， mostly use the Hession matrix sparsity and marginalization strategy in incremental equations to reduce the order of the problem. However， these methods still need a large amount of memory to explicitly construct the ultra-high dimensional Hessian matrix. Moreover， due to the sensitivity of the method to numerical changes， in order to reduce numerical errors， in actual deployment， they often rely on double precision floating point numbers to solve， which limits the application in low computing power platforms. To solve this problem， this paper proposes a visual SLAM method based on Jacobian domain null-space marginalization， which projects the landmark Jacobian matrix to its left null-space in the back-end optimization module， achieves the effect of reduction and improves the solution efficiency on the premise of avoiding the construction of Hessian matrix， and proves the equivalence of the two marginalization methods algebraically. From the perspective of numerical analysis， it is proved that the marginalization method proposed in this paper has better numerical stability and can support the single precision floating point solution， with further improvement of the efficiency. The open dataset and real vehicle test show that the method in this paper has better solution speed and accuracy than the general optimizer based on Schur complement marginalization.

Key words: automatic driving, visual SLAM, bundle adjustment, marginalization