Multi-View Point Cloud Registration Based on Improved NDT Algorithm and ODM Optimization Method

The acquisition of targets' complete point cloud model is crucial for tasks such as 3D reconstruction and disordered grasping. Shooting targets from multiple perspectives and registering point clouds from different perspectives can obtain a relatively complete point cloud model. However, small scene targets like industrial components often suffer from issues such as lack of texture and occlusion, resulting in low registration accuracy and time-consuming processes. Moreover, the registered point cloud model may exhibit uneven density and obvious noise, which hampers subsequent use and analysis of the model. To address these issues, this paper proposes an improved Normal Distribution Transform (NDT) algorithm that can automatically determine voxel size to improve registration efficiency while ensuring registration accuracy. Additionally, the Overlapping areas detection-Density uniformity-Marginal noise removal (ODM) point cloud optimization method is proposed, which first calculates the centroids of overlapping areas for density consistency processing, and then uses extracted normal features to remove noise. Our method was tested on both the Robbi dataset and the self-made dataset. The experimental results show that the improved NDT algorithm has high accuracy and efficiency for small scene targets with sizes ranging from 12.91 mm to 210 mm, and the optimized point cloud model using ODM method has uniform density and no noise.