Research On Point Cloud Completion Algorithm Based On Deep Learning

With the increasing popularity of low-cost sensors such as depth cameras and lidar,the point cloud captured by them is widely used in the field of 3D vision as a representation that can better describe the scene than 2D images.However,due to the limited resolution and mutual occlusion,the point cloud acquired by the sensor is often incomplete and sparse,and geometric information is often lost.Such incomplete scene information is insufficient for many downstream applications,such as 3D detection and autonomous driving.Therefore,recovering fine-grained full point clouds from residual point clouds has always been an important and challenging task.Due to the enhancement of computer computing power and the emergence of large synthetic shape datasets,learning data-driven prior knowledge from large datasets has become the most important research direction.Some deep learning-based models use an end-to-end method to directly infer a complete 3D shape from the original defect cloud data as input.The common problem of these methods is that they cannot generate fine target shapes.Therefore,how to reconstruct a dense,uniform and fine-grained complete point cloud from the defective cloud is a difficult point.The original point cloud scanned by the sensor inevitably contains outliers or noise,so how to reduce their impact on the point cloud completion task is also a difficulty.Aiming at the above two problems,this paper constructs a point cloud completion cascade optimization network based on feature fusion and a point cloud completion based on nonlocal neural networks with adaptive sampling,respectively.The main research contents are as follows:(1)Due to occlusion,limited sensor resolution and small viewing angle,the original point cloud collected by scanning equipment is usually sparse,irregular and incomplete,which seriously affects the effect of downstream vision tasks based on point cloud.This paper proposes a point cloud completion cascade optimization network based on feature fusion(CFF-Net),which has two branch networks.The upper branch network is used to generate coarse point clouds to extract global features with rich information.The lower branch network uses the encoder to extract point features of different resolutions,fuses them with the global features of the upper branch,passes the fused point features to the decoder through the attention mechanism,and introduces multiple loss functions to generate uniform,dense and fine-grained complete point clouds.Quantitative results show that CFF-Net reduces the chamfering distance by 20.47% compared to Top Net.Qualitative results show that the proposed CFF-Net achieves better visual performance.(2)The point cloud retains the original geometric information in 3D space.Therefore,it is the first choice for many scene understanding related applications.Unfortunately,point clouds collected by 3D sensors are usually noisy or outliers.The new completion method proposed in this paper,although it generates a uniform,dense and fine-grained complete point clouds,it is sensitive to noise or outliers.To further improve the robustness and reconstruction quality of the network,this paper proposes a point cloud completion based on nonlocal neural networks with adaptive sampling(PASF-Net).The nonlocal neural networks with adaptive sampling is used as the encoder of PASF-Net,which can reduce the influence of noise or outliers and enhance the feature extraction ability of the network.Quantitative results show that PASF-Net reduces the chamfering distance by 3.62% compared to CFF-Net.Qualitative results show that the proposed PASF-Net outputs high-quality point cloud completion models.