| Point cloud semantic segmentation is the foundation and key to achieving understanding of real-world scenes.Accurate segmentation results have vital practical significance in fields such as autonomous driving and robotics,and have a wide range of applications.In increasingly complex application scenarios,rule-based methods are no longer sufficient to meet the demands,while data-driven deep learning methods have greatly advanced point cloud semantic segmentation.However,there are some problems still exist: The model can not deeply mine the spatial information of the point neighborhood,and the aggregation method of features is also with default,which leads to the model’s inadequate understanding of local regions;The semantic information sent to the model has defects.First,the semantic information is sensitive to the directional changes of the point cloud,which easily leads to obstacles in model’s understanding.Secondly,there are mixed points of multiple categories in the boundary area.The previous model treated the boundary area uniformly without distinguishing,resulting in insufficient or even erroneous semantic feature expression in the boundary area which leading to misjudgment of model.This work focuses on the two problems mentioned above,carries out relevant research,and proposes corresponding solutions.Point cloud is a three-dimensional data form that captures the shape and structural information of the real world through three-dimensional data acquisition equipment.It has the characteristics of non-structured and irregular.In order to maximize the retention of point cloud information,this article adopts a point-based method,which can also reduce the demand for hardware resources.Based on the point-based method,corresponding solution algorithms have been designed and implemented for the mentioned problems.First,to address the problem of insufficient exploration and utilization of point cloud neighborhood information.Consider the direction and distance information between points and the points within their neighborhood range at the same time to help the model further explore information and establish an understanding of the spatial characteristics of local regions in point clouds.At the same time,a dual-space distance-based attention pooling method was designed.This method considers both Euclidean distance and feature space distance of points as important considerations,which can better preserve important features and obtain stronger expression ability of local features,thereby improving the segmentation performance.Second,to address the problem of insufficient semantic information in point clouds,it is necessary to strengthen it from two aspects: anti-directional perturbation and optimization of boundary point processing.To address the problem of semantic information being sensitive to directional changes,firstly,the coordinates of the point cloud are transformed into polar coordinates,which can make the extracted semantic information immune to rotations of the point cloud in the z-direction.Secondly,auxiliary information that is not affected by directional transformations is introduced to further weaken the impact of directional changes.To address the issue of insufficient or even erroneous semantic information extraction at boundary regions,this paper proposes to classify boundary points into two different sets and extract features from each set separately.During the feature fusion stage,learnable weights are used to adjust the contribution of the features from the two sets,in order to help the model expand the useful contributions and improve the segmentation performance. |
