Utilizing reverse engineering technology for three-dimensional reconstruction and measurement of pig body was a significant solution for cost-effective and non-contact assessment of pig body shape and condition.After compared the advantage and disadvan-tage of single and multi-view acquisition methods,we proposed in this paper a deep learning-based point cloud completion approach to restore partial pig body point cloud into complete point cloud,realizing the three-dimensional reconstruction of pig body.Our method completed the pig body point clouds collected from real production environments.This method of pig body point cloud completion was based on point proxy enhancement and layered upsampling.Initially,point proxies were generated through a combination of feature ex-traction and positional embedding.By employing Transformer for point proxy enhancement,the feature representation capabilities of these point proxies were further improved.Subsequently,based on these point proxies,a gradual and layered upsampling process was adopted to progressively restore complete point cloud with high resolution,fine granularity,and uniform distribution,transitioning from coarse level to fine level.Experiments were conducted for comparison between the proposed method and existing mainstream point cloud completion network models.Across various evaluation metrics,the method proposed in this paper had all achieved performance,which was particularly noticeable in situations where missing points in the pig body point cloud were substantial,highlighting its effec-tiveness in challenging completion scenarios.Experimental results also demonstrated the practical value of this method for completing the main parts of pig body,making it suitable for achieving three-dimensional pig body reconstruction based on partial point cloud.
关键词
猪/三维重建/深度学习/猪体点云补全/Transformer/点云上采样
Key words
pigs/3D reconstruction/deep learning/pig body point cloud completion/Transformer/point cloud upsampling
维普
万方数据