Objectives 3D sonar measurement is disturbed by the complex underwater environment and the point cloud data are usually with a high level of noise, which requires fine filtering before being applied to underwater scenes.

Methods Aiming at the shortcomings of the existing algorithms, a set of pre-processing methods for 3D sonar point cloud data is established from the aspects of reducing data complexity, local feature analysis and block filtering. First, the proposed method realizes the fine division of underwater complex space with respect to the differences in normal vector, spatial distance and echo strength of point clouds in different regions. Second, the trend surface fitting is carried out for local regions. Finally, the multi-dimensional point cloud data for error detection is constructed, the super-voxels are divided into three types of regions based on the terrain complexity, and the Grubbs test is used as a criterion of determination to realize the adaptive threshold denoising for the subregion.

Results The results show that the proposed filtering method has good accuracy for both horizontal and vertical point cloud data, with an average overall accuracy of 99.3% and an average Kappa coefficient of 0.906 for the test results.

Conclusions The results show that the synthesized filtering method has a significant improvement in accuracy compared with the traditional trend surface filtering method, and can be effectively applied to the filtering processing of three-dimensional sonar point cloud data in the underwater complex region.