Algorithms Of 3D Scene Semantic Segmentation Based On Point Clouds Data

Scene semantic segmentation is a basic problem in computer vision.For a given 2D image or 3D point clouds,the main task is to classify each pixel in the image or each point in 3D point clouds with a certain semantic segmentation algorithm and produce a predefined semantic label for it.At present,due to the success of 2D CNN technology,research on scene semantic segmentation based on 2D images has obtained fruitful results,which are widely applied in intelligent security,autonomous driving,and medical diagnostics.However,the world in which humans live is a 3D space.There are occlusion phenomena in 2D images,therefore 2D images can not completely represent the whole information of 3D scenes.Meanwhile,the image quality is easily affected by camera parameters and lighting,limiting the scope of the algorithm.While 3D point clouds preserve plentiful geometry information and they can efficiently express the information of 3D scenes.Hence,it is necessary to do 3D scene research based on 3D point clouds.This paper focuses on the problem in 3D scene semantic segmentation based on 3D point clouds.We have investigated various methods in the existing literature and find that there are rooms for optimization in solving the problem of point clouds disorder and extracting the feature relations between point clouds.This paper starts from these two aspects to study the 3D scene semantic segmentation,which mainly includes the following two parts:(1)The problem of point clouds disorder is usuaslly solved by using the symmetric function,the Max-pooling.However,the computation mechanism of the Max-pooling function is to preserve the element with the largest eigenvalue while discarding other elements,which results in feature information loss in the networks.To this end,this paper proposes a feature fusion algorithm based on attention mechanism where the point clouds features input to the Max-pooling function and the features output from the Max-pooling function are fused based on attention mechanism.Not only can this fusion method compensate for the information loss caused by the Max-pooling function but also make point clouds features contain rich high-level semantic information,which can significantly improve the accuracy of semantic segmentation algorithm.(2)There are feature relations between point clouds,which contain plentiful space geometry information and they can be viewed as an important supplement of point clouds features.In order to extract these feature relations,this paper proposes a semantic segmentation algorithm based on relation reasoning between point clouds.The algorithm constructs a local region by using 6)-NN and extract center point features in the local region by conducting graph convolutional networks.Then,the features of the center point are combined into a feature sequence and sent to a relation reasoning network composed of a bidirectional long short-term memory.The relation reasoning network reasons about the relations between points,which can achieve the purpose of extracting the feature relations between points.The experimental results show that these feature relations can largely improve the performance of the models.