This paper presents a point cloud completion algorithm using GAN to reconstruct missing data in motorcycle point clouds. By addressing viewpoint constraints, occlusion, and sensor noise, the research tackles 3D reconstruction issues effectively. A point cloud is a collection of points that form the surface of a 3D object. It is usually created by a LiDAR or depth camera. However, data obtained is often incomplete, which creates serious difficulties in restoring shapes. The GAN framework of the proposed algorithm has a perceptual encoder; a multi-stage decoder; a downsampling module; and a supervised learning module that ensure accurate completion from global structure to local detail. The perceptual encoder utilizes a tree structure and an attention mechanism to extract features with multiple dimensions. The multi-stage decoder restores the point cloud resolution progressively and the downsampling module achieves precise point correspondence matching. Test results on the ShapeNet and MVP datasets of the authored algorithm using Chamfer Distance (CD) and Unidirectional Hausdorff Distance (UHD) shows that it outperforms PointNet and Point Completion Network (PCN) for motorcycle point cloud completion, producing point clouds with reasonable structures and clear details. In addition, the point cloud completion of physical motorcycles experiment shows that the proposed method also performs good when sparsely occluded and generates models with sharp edges and uniform structures. This work proposes two novel mechanisms, multi-stage decoding and dynamic matching, for point cloud completion.
