Terrestrial Laser Scanner Point Cloud Denoising by Range Image Nlmeans Processing for Small-sized Objects

In some applications, long range operations Terrestrial Laser Scanners (TLS) achieve insufficient precision: for instance, a decorated façade will not be modeled correctly with the point cloud obtained by TLS. It may be necessary, in such cases, to combine different acquisition techniques to deal with the large size of the object to model and its fine details. Hence, data denoising in the context of TLS point clouds remains an issue: if one is capable of increasing the precision of the point cloud, it makes the instrument suitable for applications it has not been designed for. The work described in this paper deals with TLS data denoising in the context of small details like in close range applications. Compared to other denoising methods described in the literature, the herein described method puts the problem back to the 2D world. Indeed, the natural product of any TLS is a Range Image, the range being a function of two angles, vertical and horizontal. In the proposed method, each acquired station is processed by denoising this 2D function. Then, after denoising, the registration process is applied to obtain the final 3D point cloud. Two well-known image denoising methods are tested: wavelet analysis and NL-means algorithm. The presented results show that the latter method achieves good results: the standard deviation is divided by two without any increase of the noise on particular points. The method which modifies slightly the standard processing chain and which inherits the algorithm complexity of the classical 2D image processing schemes allows to extend the range of applications of TLS to smaller and fine detailed objects.