Noisy Point Clouds Registration Using FFT Based on Multi-Stage Noise Removal

Abstract

In this paper, we introduce a multi-stage fine registration technique for registering noisy point clouds. At each stage, discrete surfaces that overlap each other are simultaneously transformed into a frequency domain by a fast Fourier transform (FFT) algorithm. In the frequency domain, an adjustable function is used as the low-pass filter, and then discrete surfaces are reconstructed by an inverse Fourier transform. The iterative closest point algorithm is used to register the newly generated surfaces and obtain the registration parameters. We then registered the original point clouds by using these parameters. The next stages are implemented in the same way as in the above; only the parameters are changed in the filter. After a few stages, our method can give a better result for the registration of noisy point clouds. We experimented with the proposed method for registering many types of noisy point clouds such as noisy point clouds with different noise levels or noisy and sparse point sets.