Mixed Normal Vector Estimation Strategy for Unstructured Point Clouds

The normal vector is a basic attribute of point clouds and it has important applications in point clouds matching, surface reconstruction, feature line extraction and many other domains. Traditional normal vector estimation methods have poor robustness and are vulnerable to complex features, noise and outliers. Recently, convolutional neural network (CNN) has made great progress in normal vector estimation. However, the algorithm needs high hardware configuration and the efficiency becomes low on devices without Graphic Processing Unit (GPU). To address the problem, we propose a normal vector estimation algorithm combining the principal component analysis (PCA) and CNN. Firstly, with Surface Variation (SV) applied, the point clouds is divided into two subsets: the feature region and the flat region. Then, the PCA algorithm is adopted in the flat region while the CNN method is adopted in the region close to complex features. Meanwhile, the structure of an existing CNN is optimized to reduce the running time and improve the estimation accuracy. Experiments show that our method achieves good results in dealing with complex features and noise. Also, compared with CNN with the original network structure, the algorithm effectively reduces the computation time and improves the accuracy of normal vector estimation.