In the real world, point cloud data collection methods include LiDAR, binocular cameras, and depth cameras, but due to factors such as device resolution and surrounding environment during robot collection, the collected point cloud data is usually incomplete. In order to solve the problem of missing object shape, this paper proposed a 3D point cloud object shape auto-completion network that learnt local neighborhood information. The network consisted of a point cloud feature extraction network and a point cloud generation network. It took missing point cloud shape as input, missing part of the point cloud shape as output, and merged the input and output point cloud shape to complete the shape of the object. The experimental results show that on the ShapeNet dataset, the average chamfer distance and the average earth mover's distance are smaller than the value of the multi-layer perceptron feature extraction network and PCN network, and the two values are 0.000 84 and 0.028, respectively. Complementing the point cloud data scanned in reality also achieves the expected results, this paper indicates that the network has strong generalization and can repair different types of objects.

[2021-11-22] Accepted Paper /doc/2021.09.0383/v1

[2022-05-05] Printed Article /doc/2021.09.0383/v2