Discrete Point Cloud Registration using the 3D Normal Distribution Transformation based Newton Iteration

Abstract

The technology of three-dimensional reconstruction based on visual sensor has become an important research aspect. Based on Newton iteration algorithm, the improved 3D normal distribution transformation algorithm” (NI-3DNDT) is put forward, aiming to fix the problem of discrete point cloud registration algorithm in poor astringency and being open to local optimum. The discrete 3d point cloud adopts one order and two order derivative of piecewise smooth functions on surface, divides the point cloud space into Cubic grids, and calculate corresponding value of the mean and covariance matrix. To downgrade algorithm complexity, the Gauss function approximation of the log likelihood function is introduced, the probability density function parameters of 3D normal distribution transformation algorithm is simplified, and the Hessian matrix and gradient vector is solved through translation, rotation relation and Jacobean matrix; to make sure algorithm is converged to one certain point after a small number of iterations, it proposes that Newton iterative algorithm step be improved by employing better line search. Finally, the algorithm is put on simulation experiment and compared with other ways, the result of which proves that the suggested algorithm is able to achieve better registration effect, and Improve accuracy and efficiency