Research On 3D Point Cloud Semantic Segmentation Algorithm Based On Kernel Point Convolution

In the fields of robot system and automatic driving,environment perception is a very basic and key issue.As a key link of environment perception,3D point cloud semantic segmentation task is crucial for the system to predict the future state,avoid collision and plan the path.However,the existing research methods have many problems in the process of point cloud processing,such as information loss,incomplete feature extraction and low efficiency.Therefore,this paper proposes a 3D point cloud semantic segmentation algorithm based on kernel point convolution(KPConv)and Transformer to help the system better perceive the external environment.The specific research contents are as follows:First,in order to extract point cloud features more efficiently,a point cloud semantic segmentation algorithm based on KPConv and self-attention mechanism is proposed.In order to retain more original information of the point cloud,the KPConv operator,which directly processes the points,combined with the residual structure,can be used to perform continuous convolution operations on the input point cloud using any number of kernel points.Set up multi-layer convolution and construct the algorithm network according to the U-Net encoder-decoder mode,which can realize the dual tasks of point cloud classification and segmentation.In order to further improve the accuracy of the algorithm,a parallel spatial and channel attention module is set after the final output of the encoder convolution layer,which can selectively aggregate the context information,obtain the global receptive field,and obtain more easily recognizable features.Experiments are conducted on the Model Net40 point cloud classification dataset and the S3 DIS point cloud segmentation dataset to verify the effectiveness of the proposed algorithm model.Secondly,aiming at KPConv’s inability to overcome the variable density and uneven distribution of point clouds in complex scenes,a semantic segmentation algorithm of point clouds based on dual-path feature extraction network is proposed.The encoder part uses KPConv and Flattening Point Convolution(FPC)to design two paths to construct the convolution layer respectively.The introduction of FPC operator can make up for the shortcomings of KPConv operator in dealing with the variable density of point cloud,and can help the network extract features more robustly and accurately.Both convolution operator modules belong to lightweight networks,which are suitable for combination with each other without increasing the computational cost.The design of Dual Path Augmentation and residual feature fusion module can enable the network to learn more abundant local information,pay attention to more sufficient context information as soon as possible,and improve network efficiency.Further combining with the self-attention mechanism to aggregate global feature information effectively.The experimental results on the Model Net40 point cloud classification dataset and the Semantic KITTI point cloud segmentation dataset show that the proposed algorithm has higher classification and segmentation accuracy than other algorithms in complex large-scale outdoor scenes.Finally,in view of the problem that the learning process has high complexity and is difficult to perceive complex scenes when the network is completely dependent on point cloud convolution architecture,a point cloud semantic segmentation algorithm based on KPConv and multi-scale point cloud Transformer is proposed.Obtain three paths with different resolutions through the Farthest Point Sampling algorithm,and perform feature coding respectively.Apply KPConv to design the point embedding module to initially map and aggregate neighborhood features.The point cloud Transformer is used to construct the multi-layer network on the three paths,and the final output features are fused by the cross scale attention module,which can aggregate the dependence of mutual learning between different scales.Experiments on the Semantic KITTI point cloud segmentation dataset show that the proposed algorithm has more accurate segmentation accuracy while ensuring efficiency,and can help the system perceive the complex external environment well.